Preserve chunk_start when chunk headers arrive before payload

[AryanRG]

When a chunk header arrives without any payload bytes, ChunkedReader consumes the header and returns NEED_DATA. Its local chunk_start flag is then lost, so the first payload event has chunk_start=False when the bytes arrive later. This affects both requests and responses and violates the documented chunk-boundary contract.

For example, receiving 5\r\n, polling, then receiving hello currently produces Data(..., chunk_start=False, chunk_end=True). Receiving those bytes together produces chunk_start=True.

Keep the start marker in the reader until it emits the first payload event. Subsequent payload events clear it as before. This leaves framing, syntax validation and payload bytes unchanged.

Four regression cases cover both peer roles, ordinary and extended chunk headers delivered byte by byte, repeated polls without payload, a split payload, and a subsequent complete chunk. All four fail on upstream and pass with the fix. Includes a numbered towncrier release note.

Validation on CPython 3.12.15:

• Full suite: 82 passed (upstream baseline: 78 passed).
• Changed reader: 100% statement and branch coverage. Overall coverage remains at the upstream baseline of 98.2%.
• Black 23.3.0, isort 5.12.0, mypy 1.8.0 and git diff --check: passed.
• Additional validation checked 789 fragmentation schedules, including split chunk footers and trailers.

AI assistance was used to prepare the implementation and regression tests.