fix: output router capture wrong router logits in qwen moe models

[ITcarrot]

What does this PR do?

This PR fixes a bug in the router implementation of several MoE models (Qwen Moe like models, Olmoe, FlexOlmo).

Previously, the raw router_logits were being overwritten by the result of the softmax operation:

router_logits = F.linear(hidden_states, self.weight)
router_logits = torch.nn.functional.softmax(router_logits, dtype=torch.float, dim=-1)

This behavior destroys the raw logits information, which is often required for:

1. Casting returned router_top_value back to input type in router_top_value = router_top_value.to(router_logits.dtype)
2. Calculating auxiliary losses (e.g., load balancing loss) which typically operate on raw logits.
3. Returning correct router_logits in the model output if requested.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@Cyrilvallez @IlyasMoutawwakil

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: flex_olmo, olmoe, qwen2_moe, qwen3_moe, qwen3_next, qwen3_omni_moe

[vasqu]

Good catch yes, the logits should not go through softmax 😅

Could we have a small test added to qwen2 moe (since everyone else inherits from this) + are there any more such cases?