test(qwen3.5): preserve MoE expert stack ordering

[ukint-vs]

Summary

• keep this PR as a focused test-only follow-up now that the loader/sanitizer implementation landed through fix(qwen3.5-moe): sanitize toward the gate/up/down expert names SparseMoeBlock actually loads #671
• remove the redundant ignored output-key-only test from src/models/qwen3_5_tests.rs
• add a sanitizer unit test that verifies per-expert gate_proj, up_proj, and down_proj tensors are stacked into the current switch_mlp.*_proj loader names while preserving expert order and projection values

Audit result

The original implementation delta is already present on current main via #671, which closed #670. This branch now leaves the main qwen3_5::sanitize_weights implementation unchanged and only contributes additional regression coverage aligned with the current SwitchGLU/SparseMoeBlock naming convention.

Validation

• cargo fmt --all -- --check
• cargo test qwen3_5::sanitize_tests -- --nocapture

Refs #670 and #671

[cla-assistant]

All committers have signed the CLA.

[cla-assistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[inureyes]

Thanks for the contribution and for tracking down the Qwen3.5 MoE switch projection naming mismatch. I audited the PR against current main and the related fix that landed in #671.

The main loader/sanitizer implementation is now already equivalent on main: src/models/qwen3_5.rs has no remaining delta for this PR after #671, and #670 is closed. The useful part still worth preserving here is the extra test coverage, but it should be adjusted to the current implementation style so it complements the stronger load-through regression test that #671 added.

I will push a small follow-up commit to this PR that drops the now-redundant implementation delta and keeps only a cleaned-up, focused test. After that I will rerun the relevant qwen3_5 sanitizer regression and merge if the branch remains clean.