fix(extensions): harden plugin rpc bridge against stdin flush hangs, watchdog deadlock delays, and manifest request timeouts

[ZhuchkaTriplesix]

Problem Statement

During the extension driver resilience audit, three edge-case failure modes were identified in JsonRpcStdioClient and PluginRpcBridge:

1. Unbounded _stdin.flush(): In JsonRpcStdioClient.sendRequest, await _stdin.flush() has no timeout. If a child process deadlocks or stops draining stdin, when the OS pipe buffer (~64KB) fills up, _stdin.flush() hangs indefinitely without triggering request timeout.
2. Watchdog Deadlock Delay: When SandboxWatchdog detects a ping timeout (deadlock), it kills the handle, but cancels _exitSub beforehand so PluginRpcBridge does not receive the exit event to abort in-flight requests immediately with PluginDeadlockException.
3. Short Request Timeout vs Manifest: Extensions like clickhouse-query-ext specify analytical timeouts (e.g. timeout_seconds: 600), but PluginRpcBridge hardcodes requestTimeout = const Duration(seconds: 30).
4. Handshake Timeout Cushion: A strict 3s handshake timeout can cause false positives during cold start or bwrap setup on slow systems; also does not include stderr lines when timing out.

Proposed Solution

1. Wrap _stdin.flush() with a bounded timeout (Duration(seconds: 3)) in JsonRpcStdioClient.
2. In PluginRpcBridge.onStopped, if reason is deadlock, immediately call _failPending(PluginDeadlockException(...)) and dispose.
3. Configure requestTimeout in PluginRpcBridge based on manifest.sandbox?.permissions?.resources?.timeoutSeconds when available.
4. Increase default handshake timeout to 6 seconds and capture last stderr lines on handshake timeout.