fix(deno): wake idle serve loops on demand

[momics]

Summary

• retain the bounded native serve queue and synchronous iroh_http_try_next_request drain required by bug(deno): unknown handle / sendChunk failed under concurrent load — only the Deno adapter is affected #122
• replace the MessageChannel idle loop with a module-lifetime Deno.UnsafeCallback.threadSafe() notification carrying only an opaque serve-generation token
• unref the callback so idle servers allow process exit, while preserving shutdown, restart routing, handler draining, connection events, and fail-closed queue behavior
• correct the Deno FFI architecture documentation and stale callback constraints

Regression evidence

The regression commit deterministically observes synchronous queue drains. On main, an idle server increased from 8,773 to 36,766 polls over a 50 ms idle interval. With the fix, the count remains unchanged until Rust sends a request-ready notification.

Release-build macOS/Deno 2.9.3 measurement:

Path	User CPU	Wall time
main, idle node without serve()	0.18s	5.86s
main, idle node with serve()	5.47s	5.12s
this branch, idle node with serve()	0.16s	5.70s

Test coverage

• npm run ci
• Deno adapter regression suite (13 tests)
• 32-stream burst × 5 restart iterations
• 32-stream stop/restart routing regression
• cargo test -p iroh-http-deno
• release-build idle CPU measurement

Closes #397