Batching enhancements

[lbluque]

Summary

This branch extends the Ray Serve batched inference stack with runtime autobatch probing, broader concurrency backend support, and InferenceBatcher usability tweaks. It aligns the earlier batching-enhancements work with main's refactored batch_server layout, makes OOM splitting opt-in, and adds a configure-at-runtime API.

Main changes

src/fairchem/core/calculate/_batch.py — InferenceBatcher

• Add RayActorPoolExecutor and extend _get_concurrency_backend to support threads, processes, and ray-actors. Executor choice is now documented per workload type.
• New auto_configure_batching(probe_data, config) that runs probing and applies the result, plus exposed update_checkpoint, delete, shutdown lifecycle matching main.
• split_oom_batch defaults to False, with updated docstrings explaining heterogeneous vs homogeneous trade-offs.
• Expanded class and method docstrings with autobatch usage example.

src/fairchem/core/components/batch_server.py

• Change default split_oom_batch=False on BatchConfig, BatchPredictServer, and MultiplexedBatchPredictServer; improve OOM error messages with guidance.
• Add BatchPredictServerMixin.configure_batching(max_batch_size, batch_wait_timeout_s) for runtime tuning.
• New autobatching API:
  • AutobatchConfig and AutobatchResult dataclasses
  • probe_optimal_batch_size(predict_unit, probe_data, config) — warmup then binary-search GPU memory to find max safe atom batch size, back off by factor, derive timeout from median latency. CPU path returns conservative defaults.
  • _expand_probe_data helper to repeat probe structures to target atom counts.

tests/core/calculate/test_batcher.py

• Parametrize initialization test for new processes and ray-actors backends.
• Add test_batched_atomic_data_predictions for direct AtomicData predict path.
• Add test_autobatch_config_initialization exercising auto_configure on GPU fixture.
• Add test_probe_optimal_batch_size_cpu verifying CPU fallback behavior.

Out of scope for the committed diff: local untracked examples/batching_inference_example.py and uncommitted config edits in configs/uma/...