fix: make session mutations atomic - #4212

This pull request resolves the session mutation races reported in #4202 and the related cross-backend problems tracked in #4205. It supersedes and broadens the narrower SQLite-focused approach proposed in #4203.

Built-in session mutations previously had backend-specific gaps around concurrent appends, destructive reads, clear boundaries, cancellation, and post-commit cleanup. Those gaps could expose partial batches, return the same popped item more than once, retry an already committed mutation, or allow cancellation to interrupt resource settlement.

This change establishes consistent mutation semantics across SQLite, AsyncSQLite, AdvancedSQLite, SQLAlchemy, MongoDB, and Redis:

• Persist each logical item batch atomically.
• Claim destructive reads at most once.
• Establish one authoritative clear boundary.
• Wait for mutation and owned cleanup outcomes before propagating caller cancellation.
• Preserve Python 3.10 cancellation details.
• Use primary-authoritative MongoDB generation metadata and atomic tail claims.
• Distinguish explicit Redis transaction conflicts from ambiguous acknowledgements.
• Preserve Redis connection ownership across redis-py 7.x through 8.1 cleanup, reconnect, and reborrow paths.
• Retain backward compatibility for public call shapes, configured storage names, legacy readable data, and injected-resource ownership.

Dapr is intentionally excluded because generic Dapr state stores cannot guarantee an atomic conditional creation of an absent aggregate key with the current representation. Supporting it safely requires a separate distributed-lock contract or persisted-representation redesign.