[rpc] Fix RequestChannel backpressure race

[fxbing]

Purpose

Concurrent queue updates can pass the pause or resume pre-check before another thread changes the backpressure state. The stale operation can then acquire the lock and leave autoRead inconsistent with the current queue size.

Brief change log

• Reconcile queue size and backpressure state under the existing lock when a threshold transition may be needed.
• Preserve the lock-free fast path and log the queue-size snapshot used for each transition.
• Normalize zero thresholds and add deterministic regression coverage for stale pause and resume operations.

Tests

• ./mvnw -pl fluss-rpc -Dtest=RequestChannelTest test — 5 tests passed.
• ./mvnw -pl fluss-rpc verify — 76 unit tests and 12 integration tests passed.
• ./mvnw -pl fluss-rpc spotless:check and git diff --check.

API and Format

No public API, configuration, RPC, or storage-format changes.

Documentation

No documentation changes; this is an internal backpressure concurrency fix.

Generative AI disclosure

• Yes — OpenAI Codex