fix: preserve WebSocket send queue byte accounting during flush (#3688)

## Summary

Keep queued and in-flight messages charged to the existing byte limit
until each send succeeds. Serialize overlapping flushes, preserve FIFO
restoration on errors or cancellation, and keep concurrent drains from
releasing pending-message capacity.

The change is confined to the shared SendQueue helper and focused
queue/reconnect tests. It preserves ordinary enqueue/backpressure
behavior and does not change generated connection APIs, logging,
dependencies, or Python support. This PR targets public main
independently; no stacking dependency.

## Validation

- 34 focused synchronous/asynchronous queue and mocked reconnect tests
passed, covering all six GA/beta Responses and Realtime connection
classes.
- Repository-wide Ruff, mypy (1,562 source files), and pinned Pyright
passed; changed-file formatting and git diff checks passed.
- Separate read-only Codex review round 1 was clean (no actionable P0–P2
findings), including 20 repeated focused-suite runs.
- Local environment used the exact committed frozen lockfile because the
public-main lock freshness check requested an update; no dependency
files changed.

## Generation follow-up

SDK CODEOWNER review requested. Carry the shared-helper accounting
behavior and regression coverage into the corresponding Castiron
source/template so regeneration preserves it; no generated resource
patch is needed here.