fix(autobatching): recognise Warp OOM in memory estimation and free cache before the real run

[niklashoelter]

Problem

InFlightAutoBatcher / BinningAutoBatcher estimate a safe batch size by replaying the model forward on geometrically-growing batches until it OOMs, then backing off. Two defects on that path make models whose neighbor lists run on Warp / nvalchemiops (e.g. ORB v3) crash instead of backing off:

1. The Warp OOM isn't recognised. determine_max_batch_size only matched the substring "CUDA out of memory". Warp's allocator raises "Failed to allocate <n> bytes on device '…'", which doesn't match, so the OOM propagates and kills the run during memory estimation. In addition, the re-raise sat inside the for msg in oom_error_message loop, so passing a broadened list of signatures never worked — the first non-matching entry re-raised before later ones were checked.

2. PyTorch's caching allocator starves the external Warp pool after estimation. Once (1) is fixed, estimation survives but the crash simply moves to the first real optimization forward pass, failing on a tiny (~2.5 KB) Warp allocation. The estimation forwards leave PyTorch's caching allocator holding almost all of device memory, and Warp allocates from a separate cudaMallocAsync pool that then has nothing left.

Concretely, ORB v3 conservative on a 4272-molecule batch reproducibly died with RuntimeError: Failed to allocate 2556 bytes on device 'cuda:N'.

Fix (torch_sim/autobatching.py)

• Add DEFAULT_OOM_ERROR_MESSAGES = ("CUDA out of memory", "Failed to allocate") and use it as the default oom_error_message for determine_max_batch_size, InFlightAutoBatcher, and BinningAutoBatcher.
• Re-raise only after checking all signatures (if any(msg in exc_str …): return safe then raise), so a multi-signature list actually works.
• Release memory at the end of estimate_max_memory_scaler (shared by both batchers) with torch.cuda.synchronize() + torch.cuda.empty_cache(), guarded by torch.cuda.is_available(), so the real run isn't starved.

Tests

Adds regression tests in tests/test_autobatching.py: both PyTorch and Warp OOM wordings are recognised, and a genuine non-OOM error still propagates. pytest tests/test_autobatching.py → 38 passed; ruff clean.

Validation

ORB v3 conservative over a 4272-structure library on a dedicated GPU: 4/4 full runs complete 4272/4272 with the fix; the same code with the memory release disabled crashes with Failed to allocate 2556 bytes.