fix(mirror): recognize Kimi K3 expert shards

[terrizoaguimor]

Summary

• teach the partial-mirror planner Kimi K3's block_sparse_moe expert naming scheme
• count packed weights and scales for w1, w2, and w3 without double-counting
• use w1.weight_packed as K3's replica anchor while preserving GLM's gate_proj.weight behavior
• cover both K3 prefixes and reject near-match tensors that are not routed experts

Why

coli mirror plan currently recognizes only GLM-style mlp.experts.*.{gate,up,down}_proj.weight tensors. A Kimi K3 checkpoint therefore produces no usage-ranked shards even with valid .coli_usage history. This is the planner-only first step requested in #812; the K3 runtime still does not consume a mirror, and this PR intentionally does not close the issue.

Testing

• PYTHONDONTWRITEBYTECODE=1 python3 -m unittest -v tests.test_mirror_plan
• make check (473 Python tests passed, 59 skipped, plus the C suite)
• git diff --check

Refs #812