refactor(spec-decode): simplify Llama Eagle3 attention path for #217 (1/3)

[rjzhb]

Summary

First PR in a series refactoring the spec-decode attention path introduced in #217.

• Lift fused KV pre-write into base llama.LlamaAttention._attn (= pure perf win for target Llama when backend supports it).
• Eagle3 LlamaAttention collapses the 60-line monolith forward + 4 _maybe_* helpers + 80-line __init__ copy into a 2-method subclass: _attn overrides the draft-first-step dispatch B (correction + q-slice + DECODE / post-slice), inactive steps delegate to super()._attn.
• Correction (= spec_num_tokens - accept_lengths trim of draft_seq_lens_buf) moves from the drafter and the standalone _maybe_trim_cache_seqlens helper into a single _apply_correction method, called atomically alongside the q-slice in the active branch. Drafter plumbs accept_lengths / draft_seq_lens_buf via ForwardContext.
• Drops the temporary pre_attention_trim class flag + drafter-side isinstance gate now that the trim lives next to its only consumer.

ctx.draft_first_step_reduce stays for now (= Qwen NextN / DeepSeek V3 NextN / comm_manager still read it); follow-up PRs in the series migrate those families and then delete the field.

no behaviour change for target Llama or Eagle3 draft. Builds on #217.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bad01b8ffe

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: e06b280564

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]

temporal if-condition will be removed when 3/3 refactoring PRs are all in.