Optimize ghstack submit latency via parallelization and reduced IO

[ezyang]

Reduce wall-clock time for ghstack submit by eliminating unnecessary
network round-trips and parallelizing independent IO operations.

Key changes:

• Targeted fetch: only fetch gh/{user}/* refs instead of all branches
• Merge 3 sequential git pushes into 1 atomic push using per-refspec + force
• Replace N serial GraphQL queries with REST GETs + local PR-to-ghnum index
• Parallelize REST GET (PR info prefetch) and PATCH (PR body updates) via ThreadPoolExecutor
• Batch-push and parallel-create new PRs instead of serial push+create per PR
• Add GHSTACK_TIMING=1 env var for per-phase timing instrumentation
• Add bench/bench_submit.py for E2E benchmarking against real GitHub
• Make FakeGitHubEndpoint thread-safe and fix asyncio event loop in worker threads

[ezyang]

Stack from ghstack (oldest at bottom):

• -> Optimize ghstack submit latency via parallelization and reduced IO #334