Fix start_tls() loses StreamReader buffer data

[fizzi01]

What do these changes do?

This PR fixes a critical issue where buffered data in StreamReader is lost when upgrading a connection to TLS using StreamWriter.start_tls() or loop.start_tls(), causing the call to hang indefinitely. The issue exists in both asyncio (see python/cpython#142352) and uvloop.

The main focus is to verify and guarantee that any data buffered before upgrading a connection to TLS is not lost and is correctly transferred. These changes are effectively a port of python/cpython#142354 with absolutely no difference, except for the test that creates a TCP server, establishes a connection, buffers data, performs a TLS upgrade, and verifies that all messages (both pre-upgrade and post-upgrade) are received correctly.

Are there changes in behavior for the user?

No.

Is it a substantial burden for the maintainers to support this?

Nope.

Related issue number

Fixes #102

Checklist

• I think the code is well written
• Unit tests for the changes exist
• Documentation reflects the changes

[Vizonex]

@fizzi01 If the unittests fail know that it's caused by a rounding bug in uvloop and winloop and possibly libuv due to how it does math for getting timeouts.

[Vizonex]

@fizzi01 Thanks for the pull request btw. We don't get contributors here very often. I'll probably wind up optimizing this code in a later update since I've been trying to optimize the pure python portions for better performance.