Significantly speed up applying GTFS-RT updates

[vkrause]

With the national feeds from CH, NL and NO the time to process the updates goes down from about a minute to a second here. The cost of this was dominated by vecvec::push_back in rt_timetable::add_rt_transport.

Before:

After:

[felixguendling]

To prevent bad memory accesses, I think it would make sense to call operator[](index) as a replacement for the resize calls. The code (routing, etc.) accesses the rt_timetable as const, so the const member function is called:

https://github.com/felixguendling/cista/blob/a46b7d806b6766175af2da3a45c4cc3f78a2d2c1/include/cista/containers/mutable_fws_multimap.h#L387-L397

This has an assert for debug mode but will crash in release mode. The rest of the code assumes that the vector has n_locations entries. As probably not all locations are touched during the real-time update, the update needs to call operator[](n_locations) instead of resize to prevent invalid memory accesses in routing where the non-resizing const member is used.

It probably makes sense to give mutable_fws_multimap the same interface as vecvec (so not mirroring a std::multimap but rather a std::vector<std::vector<T>>).

[felixguendling]

rtt.location_rt_transports_[tt.n_locations() - 1U];  // resize for later memory accesses