Fix SCE mask shape on early exits

[SarveshaKumarKS]

Summary

• return SCE masks with the parameter tensor shape on every early-exit path
• cover zero and full density, small densities where k == 0, and the merge-level output shape

Root cause

The normal sce_mask path builds its mask from the variance tensor, whose shape excludes the leading model dimension. The early exits instead used zeros_like(tvs) or ones_like(tvs), retaining that dimension. sce_merge then unsqueezed the mask, and PyTorch broadcasting silently introduced an extra leading dimension into the merged tensor.

Impact

Degenerate select_topk values such as 0, as well as small positive values that round down to k == 0, now return the unchanged base tensor with its original rank instead of a rank-corrupted tensor.

Validation

• python -m pytest tests/ -q — 94 passed
• python -m pytest tests/test_sce.py -q — 6 passed
• python -m pytest tests/test_basic_merges.py::TestBasicMerges::test_sce_merge -q — 1 passed
• pre-commit run --files mergekit/merge_methods/sce.py tests/test_sce.py — all hooks passed

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[SarveshaKumarKS]

I have read the CLA Document and I hereby sign the CLA