Fix(spec decode): catch up trim bug

[rjzhb]

Summary

fix #217 that broke spec-decode catch-up on Qwen NextN and DeepSeek V3 NextN:

• catch-up seq_lens trim in drafter/eagle.py
• Llama Eagle3 pre-attention q-slice (dispatch B) in llama_eagle3.py

The trim assumed q was sliced to 1 row before the kernel. Llama Eagle3 had the matching q-slice; Qwen NextN and DeepSeek V3 NextN didn't — trim alone left q_len > cache_seqlens, giving negative kernel positions, all-masked softmax, NaN, and accept_rate=0.

Solution is to add a temp flag: pre_attention_trim.

NOTICE: this flag would be removed in the future refactor PR once adding pre_attention_trim feature to both Qwen NextN and DeepSeek V3 NextN.

Test Plan

• Qwen3.5-35B-A3B-NVFP4 + MTP local serve: no NaN in catch-up attention output, accept_rate > 0 restored.
• Llama Eagle3 (MiniMax-M2.5-NVFP4 target + thoughtworks/MiniMax-M2.5-Eagle3 draft): serves without regression.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c87743f7ec

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0c7d1ef3b5

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[LorrinWWW]

This is a correctness fix for spec. And we will have a follow-up refactoring to reduce / remove the extra if-else branching.