Correctness: opt-in deterministic conv, pruning empty-mask guard, TensorField backprop regression test

[alpsaur]

Three correctness items in the Python layer plus regression tests. No src/*.cu changes; existing built backend is unchanged.

1. Opt-in deterministic convolution (upstream #554, #504)

Convolution output depends on the input coordinate ordering: atomicAdd accumulation and kernel-map dispatch order mean that running the same conv on the same point set in a different row order yields low-order-bit differences, worst at stride > 1 and transpose. The community-confirmed workaround is to sort coordinates before the conv.

• New MinkowskiEngine/utils/determinism.py:
  • sorted_coordinates(sparse_tensor) returns a new SparseTensor with rows in canonical lexicographic coordinate order. Implemented as a stable per-column radix argsort (batch column most significant) so it never packs coords into a single integer rank (overflow-free), and it is correct for arbitrary spatial dimension D and negative coordinates.
  • Process-wide set_deterministic(mode) / is_deterministic() flag (off by default).
• MinkowskiConvolution / MinkowskiConvolutionTranspose forward route the input through sorted_coordinates when the flag is on (skipped for the use_mm, kernel_volume == 1, path which is a plain matmul). Documented cost: one coordinate sort per conv; intended for reproducibility/debugging, not throughput.
• Exported as ME.set_deterministic, ME.is_deterministic, ME.sorted_coordinates.

Default behavior is unchanged (opt-in only).

2. Pruning empty-mask guard (upstream #579)

Reproduced on this fork first: an all-False mask through MinkowskiPruning on CUDA followed by a downstream conv and backward. It does not crash. The CUDA pruning kernel already emits a valid empty SparseTensor for the size-0 case, downstream conv produces zero-size outputs, and backward yields finite (empty, zero) gradients. Per the plan, no code change was needed; added a regression test that locks in the behavior.

3. TensorField backprop regression test (upstream #395)

The reported merge_sort cudaErrorIllegalAddress in inverse_mapping backward was fixed upstream by commit 02fc608, which is in this fork's history. Added a field -> sparse -> conv -> slice -> field backward test (~50k points, D=3, batch 2) run over 5 iterations with varying N to catch intermittence; asserts finite gradients. Passes (no xfail required).

Tests

New CUDA-gated modules (skip cleanly on CPU-only CI):

• tests/python/determinism.py
• tests/python/pruning_empty.py
• tests/python/tensorfield_backward.py

Verified on RTX 5090 / CUDA 12.8:

• New tests: 6 passed on GPU, 6 skipped on CPU-only.
• Core subset convolution.py (minus TestPCD), pruning.py, half_precision.py: 26 passed, 4 skipped, 2 deselected.