[kyo-zzz]

Fixes #14769.

ModelMixin.save_pretrained cleans a previous save by removing files that match the shard pattern of the current save, but it only recognized sharded files. Single-file checkpoints of the other container were left behind: re-saving a .safetensors checkpoint as .bin (or the reverse) kept the old file on disk, and because from_pretrained prefers sharded and safetensors checkpoints, the stale one silently took precedence over the freshly written weights — no error, no warning.

Relationship to #14844

#14844 fixed the sibling problem: cleanup matched another variant's shards through a bare prefix, and a formerly-sharded checkpoint left a stale index behind. It scopes the match to the container being written (filename.removesuffix(ext)), so a .safetensors → .bin re-save still leaves the old .safetensors on disk.

This PR is rebased on top of it and keeps that anchoring, then extends the cleanup to both containers.

Changes

• Extract the cleanup into _get_superseded_checkpoint_files, which returns every artifact of the current save's variant that the new save replaces: single-file weights, shards, and sharded indexes, in either container.
• Scope is the variant stem (diffusion_pytorch_model vs diffusion_pytorch_model.ema), so checkpoints of other variants that share the directory are preserved. Shards are matched through _REGEX_SHARD group 1 rather than a prefix, which keeps the fix: scope save_pretrained shard cleanup to the saved variant (avoid deleting another variant's shards) #14844 regression fixed.
• Sharded indexes are cleaned as well, which fixes the stale-index FileNotFoundError case from the issue matrix.

Tests

tests/models/test_modeling_common.py::TestModelUtils

• test_save_pretrained_preserves_other_variant_shards (from fix: scope save_pretrained shard cleanup to the saved variant (avoid deleting another variant's shards) #14844) — unchanged.
• test_save_pretrained_removes_superseded_checkpoints[None|ema] — parametrized over both variants, keeps a checkpoint of the other variant in the directory, switches the container in both directions, and asserts that the superseded file is gone, the other variant survives, and the reloaded weights match the freshly written ones.

Verified red-then-green against the current main: reverting only src/diffusers/models/modeling_utils.py to main makes both new cases fail (assert 'diffusion_pytorch_model.ema.safetensors' not in [...]) while the #14844 tests still pass; with the fix applied all four pass. ruff check and ruff format --check are clean.