fix: SendRequest hangs forever when server process dies

SendRequest's select only watched ctx.Done() and responseChan, missing
the done channel. When the server process dies (crash, pipe break,
startup failure), the reader goroutine exits but SendRequest blocks
forever waiting for a response that will never come. This causes MCP
clients to become permanently unresponsive.

Root cause: readResponses exited silently on EOF without signaling the
done channel, so in-flight requests had no way to know the server died.

Fix (5 changes in client/transport/stdio.go):
1. Add closeDone() using sync.Once to safely close the done channel
   from multiple goroutines without panicking on double-close.
2. readResponses calls closeDone() on unexpected exit (EOF/error),
   so in-flight requests unblock automatically on server death.
3. SendRequest's select includes <-c.done (both pre-check and
   response-wait) to return ErrTransportClosed immediately.
4. SendNotification gets matching done+ctx pre-check for consistency.
5. Close() uses a separate closeCleanupOnce to always perform resource
   cleanup (stdin, stderr, cmd.Wait) even when readResponses already
   called closeDone(), preventing FD leaks and zombie processes.

Includes 8 regression tests covering server death, concurrent close,
FD cleanup, and concurrent request stress scenarios.

Author: Ismar Iljazovic