Ignore explicit zeros in sparse embedding counts

[HuaTNA]

Fixes #4068.

Filter out explicit zero values after coalescing sparse embeddings in compute_count_vector. Sparse and dense representations of the same embeddings now produce the same nonzero counts, including when duplicate sparse entries cancel to zero.

Regression coverage includes 1D and batched embeddings, coalesced and uncoalesced inputs, and all-zero inputs. The four new cases failed before the fix; pytest -q tests/util/test_tensor.py passes all 23 tests afterward. Ruff lint, formatting, and whitespace checks passed.

AI assistance was used to investigate, implement, and test this patch.