[moetemp]

What was changed

This PR adds temporalio.client.Stream, which makes the Stream calls outside Workflow code, with Client.create_stream, Client.get_stream, Client.to_stream and Stream reads (Subscription).

• Client.create_stream is create-or-get by Stream id. On a Stream that exists, it ignores the options and returns that Stream, and on a closed one it raises StreamClosedError. It takes an optional backend name, so it doesn't depend on the Store.
• Client.get_stream(definition, *, id, backend=None) makes no call. Reads decode each Record with the definition's payload type.
• Client.to_stream(handle, *, payload_type=None) gives the Stream of a handle that came as data, such as a query result. It makes no call. Reads decode each Record as a JSON value, unless payload_type is given. stream.handle gives the data to pass on.
• publish starts the write at the call and returns a future. The writes of one Stream keep the order of the calls. A write error that no caller got raises at the next call of the Stream, and close closes first, then raises it.
• rpc_metadata and rpc_timeout are keyword-only on publish, subscribe, get_offset and close.
• close on a Workflow's Stream raises StreamNotOwnerError with no call, and so does async with at __aenter__.
• A read yields Records, Superseded items that withdraw earlier Records, and Closed last. It long-polls and starts at an offset, at END or at the newest last Records. It matches each item on the oneof of data, superseded and close. The long poll's loop, which carries the offset across empty pages, is here.
• A client Producer id has a key for its process. A forked child makes its own key, so two children of one parent never write as one Producer.
• A read keeps the offsets that it gave of each Producer that a later Superseded item can withdraw, as runs of consecutive offsets. A batch has consecutive offsets, so a read of a million batched Records keeps almost nothing, and a read of a million single Records keeps about 17 MB, not 44 MB.
• An in-Python fake of the crate's runtime lets the facade tests run with no Store.
• _convert.py and _writer.py are here and not in S3 and S4, since their first users are here.

Depends on S4 (#374). No changelog fragment, since S9 has the Streams entry.
Part of the Brook review series.

Why?

This is the client half of the API, which S6 runs on Redis and in memory.

Checklist

1. How was this tested:
   • poe lint, poe gen-docs and pytest tests/contrib/streams.
   • Fork CI passes, except the arm jobs, since the fork has no arm runners.