Move S3File writes onto a multipart write session #1227

Sub-issue of #1211, which records the design decisions.

Use case

S3File (also used by AioS3File) tracks its own part futures, in-flight sets, failure checks, and abort. Its lifecycle state is implicit, which lets discard() followed by commit() upload the buffered last part alone, lets a commit() after an abort failure complete the discarded upload, and keeps failed part futures in reference cycles.

Proposed change

Add the internal _S3WriteSession and S3MultipartStateError to pyathena/filesystem/s3_writer.py as specified in #1211, and move S3File onto it:

• _initiate_upload, _upload_chunk, commit, discard, and _close_without_commit delegate creation, part uploads and copies, bounded concurrency, the single-request path, completion, and abort to the session.
• multipart_upload becomes read-only; multipart_upload_parts is replaced by a snapshot of the completed parts.
• commit() after a discard or a failure raises S3MultipartStateError; a write after a failure raises before buffering.

Behavior changes are listed in the PR for the 4.2.0 release notes.

Validation plan (if implementing)

• Unit tests of the session without AWS: state transitions, admission and readiness, waiter cancellation, scheduling failures before and after a request starts, late callbacks during an abort, two abort callers, queued versus running completion, abort failure and retry, interrupts, and one-worker pools.
• GC-disabled weak-reference tests that the session releases failed request futures and part bodies after the error is consumed.
• The existing S3File and AioS3File tests, rewritten where they injected multipart_upload_parts or set multipart_upload, with both S3ThreadPoolExecutor and S3AioExecutor.
• The filesystem integration tests against the CI bucket (append, exclusive create, transactions).