Fix sparse Jacobians with empty dependencies and singleton batch axes

[Giuseppe Carleo (gcarleo)]

Sparse propagation can fail for batched attention and complex Slater wavefunctions:

• find_out_idx removes singleton position axes, but remove_zero_entries indexes the original operands. Restore the broadcast position frame before selecting nonzero entries; a size-one batch ahead of a multi-head contraction otherwise produces out-of-bounds mask indices.
• A sparse Jacobian may have no rows. Treat its largest dependency index as -1, allowing the existing zero-Jacobian densification path instead of reducing an empty array without an identity.

Regression tests compare multi-head attention values, gradients, and Laplacians against JAX autodiff for unbatched, singleton, nested-singleton, and two-item batches. The empty-mask test now also covers zero rows.

Validation with JAX 0.10.2 and two CPU devices:

JAX_NUM_CPU_DEVICES=2 python -m pytest -o addopts='' -q test --ignore=test/experimental
# 148 passed, 328 subtests passed
pre-commit run --all-files