Upgrade-tests use git to clone the current repository and
check the necessary versions out.

The cloned repository was always saved in vshard/test/var, even
when the actual --var argument for test-run.py was /tmp/var.

Lets better make the copied code also stored by the path given
in --var.

NO_DOC=internal

It makes the unknown variables treated as errors, not as 'nil's.
Otherwise it is too easy to use a wrong variable name somewhere
and get it as nil, and the tests would even pass, but not test
what they are supposed to.

NO_DOC=internal

The only map-reduce function is router.map_callrw(). At some point
there was a task to introduce a new mode of Map-Reduce - partial,
by bucket IDs, not on the whole cluster. For that task was
introduced a new function router.map_part_callrw() which had the
same Map-Reduce and error handling stages. Only Ref stage was
different.

The new helpers in this commit were supposed to reuse some code
between those two map-call functions.

Later it was decided to leave just one map-call function and add a
new option to it. But these new helpers still look useful to have
as separate functions. They make the map-call function really
small and simple.

NO_DOC=internal
NO_TEST=refactoring

It ensures the ref is still in place. A read-only operation. It is
going to be used in the future commits about partial map-reduce.
Router will be going potentially more than once to some storages
and at all times it would ensure the ref is still in place.

NO_DOC=internal

The storage-size of the Partial Map-Reduce feature.

NO_DOC=later

The previous commit was failing some tests. Lets patch them up.
That commit isn't amended so as to keep its original shape in
respect to the external contributor.

NO_DOC=bugfix

Introduce a partial ref-map-reduce API for vshard. It guarantees
that in case of success the function is executed exactly once on
the storages, that contain the given list of buckets.

NO_DOC=later

The previous commit was failing some tests. Lets patch them up.
That commit isn't amended so as to keep its original shape in
respect to the external contributor.

NO_DOC=bugfix

This is useful for RW map-reduce requests which need to send
multiple network requests in parallel to multiple masters.

In-parallel means using is_async netbox feature. But it only
works if the connection is already established.

Which means that the connection establishment ideally must also
be parallel.

NO_DOC=internal
NO_TEST=already covered

There were a number of minor issues with the previous several
commits, like the tests running way too long or some cases not
being covered or the code being non-critically suboptimal.

Lets fix them all. The original commits aren't amended so as to
keep their original shape in respect to the external contributor.

NO_DOC=bugfix

There are 2 Ref-Map-Reduce functions right now - map_callrw() and
map_part_callrw(). Their only difference is that the former refs
the whole cluster, while the latter refs only a subset of
storages. The rest is the same.

There is an idea, that better lets merge these functions into one
and make the bucket IDs an option.

The commit extracts the Ref stages of both functions into separate
helpers which will allow to keep this future single function very
short and simple.

NO_DOC=internal
NO_TEST=refactoring

The behavior is regulated with the new bucket_ids option.

@TarantoolBot document
Title: vshard: `bucket_ids` option for `router.map_callrw()`

The option is an array of numeric bucket IDs. When specified, the
Ref-Map-Reduce is only performed on the masters having at least
one of these buckets. By default all the stages are done on all
masters in the cluster.

Example:
```Lua
-- Assume buckets 1, 2, 3 cover replicasets UUID_A and UUID_B.
res, err = vshard.router.map_callrw(func, args, {bucket_ids = {1, 2, 3}})
assert(res[UUID_A] == {func_result_from_A})
assert(res[UUID_B] == {func_result_from_B})
```

Lets merge all map_callrw() tests into a single file.

NO_DOC=internal

When there were only 2, all cases would either cover a single
replicaset or "all" of them. Lets make them 3, so that some tests
actually cover a part of a cluster which is not just a single
replicaset.

NO_DOC=internal

'moved_buckets' function would treat as "moved" all the buckets
which are not strictly ACTIVE. But that isn't optimal.

Also the 'moved_buckets' func would assume that when ref creation
is started, by the end of it the buckets stay unchanged. That
isn't true.

Thirdly, the moved buckets could contain the destination where
did they move to. Returning this to the router would make the
re-discovery faster.

Fourthly, PINNED buckets were not considered ACTIVE.

The commit fixes all these issues.

Firstly, when a bucket is SENDING, returning an error right away
isn't good. The router would just keep retrying then, without any
progress. The bucket is actually here, it is not moved yet.

Better let the storage try to take a ref. Then one of 2 results
are possible:
- It waits without useless active retries. And then SENDING fails
    and becomes ACTIVE. Ref is taken, all good.
- It waits without useless active retries. SENDING turns into
    SENT. Ref is taken for the other buckets, and this one is
    reported as moved.

Similar logic applies to RECEIVING.

Secondly, after a ref is taken, the not-moved buckets could become
moved. Need to re-check them before returning the ref. Luckily,
the storage can use bucket_generation to avoid this double-check
when nothing changed in _bucket.

NO_DOC=bugfix

During the partial Map-Reduce the router might visit some storages
more than once. Happens when after a ref on storage-A another
storage-B reports A as having taken some buckets.

Then router would come back to A to confirm that. The storage
still must hold its previously created ref in order for such
checks to make any sense. Otherwise any of the previously
confirmed buckets could have had escaped by now.

Without the ref-checking the router could reach the Map stage and
send some Map requests even though could detect earlier, that not
all storages would succeed.

This wasn't strictly speaking a bug, but it was clearly suboptimal
behaviour leading to the requests being executed not on all the
needed storages while the others would report errors.

NO_DOC=internal

It tests not only partial Map-Reduce. It covers a bit of the full
one as well.

NO_DOC=internal