fix(websocket): isolate routes per server without callback cycles

[justrach]

Closes #227.

Scopes native WebSocket route registries to each TurboServer/listener, carries registry identity through accepted connections and queued work, preserves app-local live replacement, and retains owned handler/guard snapshots for in-flight jobs.

Callback ownership is Python-GC-visible, so bound methods and closures that capture their app/server do not form hidden native cycles. Native registry entries borrow the state-owned callbacks; queued/in-flight work owns explicit snapshots. Interpreter checks happen before mutation and registration remains transactional.

Validation on current main:

• full Python suite: 597 passed
• combined feat(websocket): add pre-upgrade authorization and Origin guards #201/feat(websocket): support nonblocking duplex I/O with bounded backpressure #202/Fast routes drop Response headers and cached status metadata #226/WebSocket routes leak across app instances through a global native registry #227 focused matrix: 140 passed
• exact Zig 0.17 dev1862: 50/50
• same-process two-port concurrent isolation, live replacement, teardown, bound-method and capturing-closure GC: passed
• paired persistent-WebSocket performance: within the 10% gate
• exact diff: three intended files; current generated perf baseline/history retained byte-for-byte

No production deployment or release is included.

[github-actions]

Performance Regression Report

Runner: ci-pr | Duration: 5s per endpoint | Threads: 4 | Connections: 100 | Workers: 24

Endpoint	req/s	avg latency	p99 latency	threshold	status
GET /health	84,593	0.21ms	1.05ms	80,000	OK
GET /	86,354	0.23ms	1.00ms	80,000	OK
GET /json	82,695	0.27ms	0.94ms	75,000	OK
GET /users/123	82,603	0.27ms	0.84ms	75,000	OK
POST /items	47,923	0.48ms	1.20ms	45,000	OK
GET /status201	76,637	0.26ms	1.17ms	65,000	OK
AVERAGE	76,801			50,000	OK

✅ All endpoints pass regression thresholds