Remove unused sparse convolution backends

[blackencino]

Remove unused sparse convolution backends

Summary

Removes all sparse convolution backends except gather-scatter, and eliminates the
SparseConvPackInfo / ConvPackBackend C++ wrapper layer entirely. Upcoming NanoVDB
changes will provide alternative convolution methods (starting with a CUTLASS-based
backend); these legacy implementations are no longer needed and were a significant
source of compile time, code complexity, and bugs.

Net: -22,500 lines across 44 files. ~1.1 MB of CUDA/C++ deleted.

What changed

Deleted backends: ImplicitGEMM (4 variants), CUTLASS, LGGS, ME (minkowski engine
gather-scatter). 20 C++ source files removed.

Deleted SparseConvPackInfo: The monolithic C++ variant class (~20 std::optional<Tensor>
fields) is gone. Replaced by two standalone C++ functions exposed to Python:

• build_kernel_map() -- computes the gather-scatter neighbor map
• sparse_conv_kernel_map() -- runs the autograd convolution with raw tensors

Simplified ConvolutionPlan: No longer wraps a C++ object. Stores a typed Python
backend (_GatherScatterBackend, _DenseBackend, or _MatmulBackend) directly.
Backend selection collapsed from ~130 lines to ~20.

Build: CUTLASS dependency and cmake helper retained (needed for future backend).
9 .cu + 1 .cpp removed from CMakeLists.

Tests: WIP conv tests/benchmarks deleted (used removed APIs directly). Main unit
tests updated and passing.

What this PR does NOT do

• Bug fixes to gather-scatter: The existing gather-scatter algorithm is unchanged.
  Known issues (e.g. redundant .cpu().contiguous() in the hot path, weight permutation
  on every call) remain and will be addressed in a follow-up.
• Remove the halo convolution: GridBatch.sparse_conv_halo() stays in the public API.
  It should be deprecated in a future PR once we have buy-in, but that is out of scope here.
• New backends: The planned CUTLASS-based backend is not included. This PR clears
  the path for it.

[sifakis]

Looks good to me. Nice cleanup