[Feature] AsyncBatchedCollector: coordinator loop and direct submission mode

Rewrite the AsyncBatchedCollector to use a coordinator thread that
pipelines env stepping and batched inference without a global sync
barrier. Add a `direct=True` mode where each env thread submits
directly to the InferenceServer, eliminating the coordinator thread
and its serialization overhead.

Benchmark results (8 mock pixel envs, Nature-CNN, CPU):
  AsyncBatchedCollector direct:    3183 fps (+72% vs coordinator)
  AsyncBatchedCollector threading: 1850 fps (coordinator mode)
  AsyncBatchedCollector mp:        1042 fps (coordinator mode)

Co-authored-by: Cursor <cursoragent@cursor.com>
ghstack-source-id: 861a24d
Pull-Request: #3499
Co-authored-by: Cursor <cursoragent@cursor.com>

Author: vmoens