[Spec Decoding] Streamline batch expansion tensor manipulation

[njhill]

There are some inefficiencies in the spec decoding logic, particularly related to batch expansion and handling of mixed spec decode / non spec decode batches.

• Have split_batch_by_proposal_len method split the batch in a single iteration rather than iterating separately to get the spec/non-spec lists
• Call _run_no_spec rather than _run_speculative_decoding_step in the case that all sequences have max_proposal_len == 0
• Add fast-path _contract_batch_all_spec method for the (common) case that all batch sequences have spec decode enabled which excludes logic to split/recombine the spec/non-spec sequences
• Simplify the _split_scoring_output method used in _contract_batch to avoid unnecessary intermediate tensor manipulation

In an anecdotal test of mlpspeculator with bs=1 this gives a consistent 2-3% increase in throughput.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

• Comment /ready on the PR
• Add ready label to the PR
• Enable auto-merge.

🚀

[cadedaniel]

Thanks so much!