[Rust][Arrow] Coordinate explicit close with recovery

[teodordelibasic-db]

What changes are proposed in this pull request?

Arrow explicit close previously began supervisor teardown only after its flush attempt completed. If the supervisor was recovering, backoff, connection setup, ready-signal waiting, authentication invalidation, or replay could therefore continue until a separate timeout resolved that phase.

This PR coordinates close and recovery through one level-triggered close state and one supervisor-owned finalization path.

• Interrupt recovery backoff, reconnect setup, ready-signal waiting, authentication invalidation, and partial replay when close is requested.
• Keep active ACK processing alive until the close flush completes, then let the supervisor detach the sender and finalize exactly once.
• Prevent a replacement sender from being published after close teardown begins.
• Preserve an already-ready ACK, peer error, or reconnect authentication rejection instead of masking it with an older transport error or synthetic flush timeout.
• Rebuild replay state from the durable watermark and retain only the correct unacknowledged batch suffixes when close interrupts a partial replay.
• Keep cancelled and repeated close calls resumable and idempotent through one stored outcome.
• Preserve the pending-relative ACK deadline behavior introduced by [Rust][Arrow] Make ACK deadlines pending-relative #654 when replay and close coordination overlap.

Intentionally deferred

• Support for concurrent ingest and close across an FFI boundary, which remains documented as unsupported.
• Fallback finalization after an abnormal supervisor exit or panic.
• Reuse of the original close deadline after a close future is cancelled and resumed.
• Post-finalization credential invalidation timing and cleanup.
• Publication-gap micro-races outside the supported concurrency contract.
• Replace the private close-cancellation sentinel with typed internal reconnect/commit outcomes.
• Additional duplicate-waiter and header-provider-specific cancellation coverage.

No public API, FFI signature, or ABI changes are introduced.

How is this tested?

• test_close_interrupts_recovery_backoff: close interrupts a parked recovery backoff and preserves the triggering error.
• test_close_interrupts_reconnect_do_put_handshake: close cancels a blocked reconnect DoPut handshake.
• test_close_interrupts_reconnect_ready_wait: close cancels reconnect ready-signal waiting.
• test_close_preserves_reconnect_auth_rejection: a reconnect authentication rejection wins over the older transport failure and remains stable across repeated close calls.
• test_close_during_partial_replay_retains_exact_unacked_suffixes: close during partial replay retains only the durable-watermark suffix and later pending batches.
• ready_process_error_wins_over_visible_flush_completion: a ready peer error is processed before a simultaneous close-flush completion.
• ready_ack_work_is_polled_before_visible_flush_completion: ready ACK work is processed before a simultaneous close-flush completion.
• ready_reconnect_result_wins_over_ready_close_request: a completed reconnect result is not discarded when close becomes ready at the same time.
• empty_close_target_does_not_mask_terminal_error: closing an empty stream preserves its terminal error.
• explicit_close_cancellation_finalizes_before_close_publication: a private close cancellation preserves the original recovery error before close-state publication becomes visible.

Fixes #657.