Remove placeholders

[AdrienVannson]

Summary

Hello!

This is a first step in removing the PLACEHOLDER mechanism to greatly improve performance, make the whole code more Pythonic and simple and resolve a lot of bug. The motivation is more deeply described in #631

This PR completely removes placeholder values from messages. Instead, fields are always set with there default value (empty string, 0, None for optional, members of a oneof and messages, ...). Therefore, all the message field are marked as optional and can be set to None, which is coherent with the specification (see the issue).

As a consequence, this fixes the following issues:

• One-of pattern matching not supported by static type analysis tools #601 One-of pattern matching not supported by static type analysis tools
• Nested message with repeated field not in to_json when mutated after initialization #543 Nested message with repeated field not in to_json when mutated after initialization
• nested empty types not in to_dict #199 nested empty types not in to_dict
• Message with oneof called method .to_pydict/.from_pydict fails  #610 Message with oneof called method .to_pydict/.from_pydict fails
• [Feature Request] Compatibility of generated Pydantic models with FastAPI for proper OpenAPI Swagger serialization #478 Compatibility of generated Pydantic models with FastAPI for proper OpenAPI Swagger serialization (I haven't tested but probably)
• to_dict on empty nested types #200 to_dict on empty nested types (no need for this PR anymore)
• (others, I need to complete the list)

Before merging it, I think it would be good to finish the work in progress (at least my other PR and another fix of the generation of the comments), and release a new version with all the recent bugfixes.

Performance

These changes also make it possible to remove the __getattribute__ from messages. As expected, this has a great impact on the performance.

With the following messages:

message Test {
    bool x = 1;
    int32 y = 2;
    float z = 3;
}

message List {
    repeated Test msg = 1;
}

The following code takes 3.24s on my computer with the original (master) implementation with the Rust codec, and only 2s with my PR (without the Rust codec). Thus, even if the Rust codec is not supported for now, it is not a problem since the performance are better anyway.

msg = List(msg=[Test(x=bool(i % 2), y = i, z = i / 2) for i in range(100000)])
b = bytes(msg)
msg2 = List().parse(b)
assert msg == msg2

Breaking changes

This PR introduces two small breaking changes:

• Lazy-initialization of messages is no longer possible, since uninitialized message fields now have None as value and getattribute is no longer defined.
• The serialized_on_wire function is now removed, since messages that were not serialized on the wire are now represented as None values. This function didn't really make sense in the first place imo, since being serialized is a property of a field of a message, not a property of a message itself (google's version has HasField, which indeed takes the field as a parameter).

Checklist

• If code changes were made then they have been tested.
  • I have updated the documentation to reflect the changes.
• This PR fixes an issue.
• This PR is a breaking change (e.g. methods or parameters removed/renamed)

[Gobot1234]

I do like this after more thought. Really do wish that PEP 505 happened

[AdrienVannson]

Thanks! I applied your suggestions

[Gobot1234]

Is this no longer required?

[AdrienVannson]

I let them for now as a reminder to change something else, but I'll remove it soon (in the fork)

[AdrienVannson]

(done)

[AdrienVannson]

Note: I already merged this PR in the forked repo https://github.com/betterproto as discussed on Discord, so it's probably not worth it to solve the merge conflicts here