Fix equivolume batch broadcasting

[ShiroKSH]

Summary

• reshape per-mesh tetrahedron means along the batch axis
• add a regression test with different batch and tetrahedron counts
• correct the documented output shape and example values

Root cause

equivolume reduced tetrahedron volumes to a tensor of shape (B,), then reshaped it to (1, B) before subtracting it from volumes shaped (B, T). This raised a broadcasting error when B != T and mixed means between meshes when B == T.

The mean now has shape (B, 1), so each mesh is compared with its own mean volume.

Validation

• all four TestTetMeshMetrics checks pass in an isolated CPU PyTorch environment
• batched output matches per-mesh evaluation
• backward pass produces finite gradients; gradcheck passes
• flake8 and py_compile pass for the changed files

[ShiroKSH]

@Caenorst - could you take a look at the equivolume batch-broadcasting fix in this PR when you have a chance, or route it to the right reviewer? I can make any requested changes or add tests. Thanks!

[ShiroKSH]

Hi maintainers, this PR has been open for over two weeks. When you have a moment, I would appreciate a review. Thank you.

[Caenorst]

Apologies for the delayed answer. Thank you for the contribution, good catch with the issue, can you just revise the tests as suggested and I'll be happy to merge.

[Caenorst]

This is exposing a caveat in the original test_equivolume because B == T. I think we should just fix test_equivolume to have a different batch size, maybe just increment B. It's better than making new test I think.