[BugFix] correct topk_softmax/topk_sigmoid routing weights at (128, 8)

[Zymonody7]

mcoplib _moe_C.topk_softmax returns biased weights when a correction bias is passed and _moe_C.topk_sigmoid returns softmax weights, both only at the (num_experts=128, topk=8) tile (issue #270). Expert ids are correct, so recompute the weights from gating_output at the selected ids and keep the kernel untouched. The extra gather runs on gating tensors only, negligible next to expert MLPs.

Purpose

Work around #270: mcoplib _moe_C.topk_softmax returns biased weights when a
correction bias is passed, only at the (num_experts=128, topk=8) tile. Expert ids
are correct in all cases, so this patch recomputes the weights from gating_output
at the kernel-selected ids via the existing patch/bugfix mechanism. The kernel is
untouched; all other shapes take the original path; the extra gather runs on gating
tensors only.

The patch also covers topk_sigmoid for older kernels: sigmoid is broken on
mcoplib 0.4.2 (per #270) but already correct on 0.4.5; the recompute is identical to
kernel output there, so it is harmless.

Test Plan

C500, MACA 3.7.1.5, mcoplib 0.4.5+maca3.7.0.37.torch2.8, shapes
E∈{64,128,256} × K∈{4,8} × {bias, no-bias} (repro script attached in #270 thread).

Test Result

Before: E=128 K=8 softmax bias=Y: max_diff = 3.03 ← broken (all other shapes ≤1e-7)
After: E=128 K=8 softmax bias=Y: max_diff = 0.0 (renormalize=False/True both)
Sanity: all other shapes unchanged ≤1e-7.

Also confirmed not present on mcoplib 0.3.1+maca3.3.0.15 (no bias arg there).

(Optional) Documentation Update

None.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.