Release v0.5.8: correctness release (interpolation fix, deterministic conv)

[alpsaur]

Release prep for v0.5.8, aggregating PRs #15/#16/#17:

• Boundary interpolation renormalization (upstream Wrong results with MinkowskiInterpolation() NVIDIA/MinkowskiEngine#477, open since 2022)
• Opt-in ME.set_deterministic(True) (upstream Non Reproducible Outputs on GPU when using MinkowskiConvolution and stride > 1 NVIDIA/MinkowskiEngine#554)
• Regression tests for Program terminated when trying to get a kernel_map NVIDIA/MinkowskiEngine#579 (empty pruning) and merge_sort: failed to synchronize: cudaErrorIllegalAddress with TensorField during training on some ScanNet scans NVIDIA/MinkowskiEngine#395 (TensorField backprop)
• Operations cookbook (docs/operations.md)

Release audit performed on merged master (RTX 5090, CUDA 12.8, torch 2.9.1):

• Full relevant suite green: interpolation_renorm (9/9), determinism, pruning_empty, tensorfield_backward, interpolation, half_precision, convolution, pool
• End-to-end determinism check: stride-2 conv+transpose net, bitwise-equal outputs across random input permutations with the flag on; flag-off path unaffected
• One pre-existing failure surfaced and dispositioned: interpolation::test_gpu hits the same torch 2.x gradcheck IndexError as the already-skipped ::test variant; verified failing identically on the PRE-fix build (file-swap A/B), so it is unrelated to the renorm change. Added to the legacy skip list with rationale; gradient correctness is covered by the new renorm gradcheck tests.

After merge: tag v0.5.8 -> wheels publish -> ABI acceptance on hardware.

🤖 Generated with Claude Code