fix(mcp): move blocking stdio I/O off the cooperative pool (#431)

AnyMCPConnection.callTool carried a comment saying it ran blocking stdio I/O
"off the cooperative thread pool" via Task.detached. Task.detached does not do
that: it changes context inheritance -- priority, task-locals, cancellation --
not the executor. A detached task is still scheduled on Swift's cooperative
executor, so the synchronous MCPConnection.callTool kept occupying a cooperative
worker while it took an NSLock and blocked in read(2) on the subprocess.

The cooperative pool is sized to the core count, so a handful of concurrent MCP
tool calls could stall unrelated async work that had nothing to do with MCP.

Now dispatched to DispatchQueue.global(qos: .userInitiated) and bridged back
with withCheckedThrowingContinuation, which actually leaves the pool.

Verified the MCP path still works end to end after the executor change:
`apfel --mcp mcp/calculator/server.py "add 41 and 1"` -> tool executed, 42.

Diff taken from candidate PR #442.

Closes #431

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011ccLBbEaVVd4sJyUd5wyhA