[V1] [Hybrid] Enable Full CUDA Graph (decode-only) for Mamba layers

[tdoublep]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

Allow FCG to be used for batches that contain only decode requests. This PR will close the last remaining performance gap to V0 for hybrid models.

Should be merged after #21367 (diff will be much smaller)

cc @heheda12345 @tlrmchlsmth @fhl2000

Test Plan

TBD

Test Result

TBD

(Optional) Documentation Update

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request refactors the CUDA graph support in attention backends to be more granular by using an enum instead of a boolean. This is a good change that allows enabling full CUDA graph for decode-only batches in hybrid models, such as those using Mamba layers. The changes are well-implemented across various backend files. I've found a minor copy-paste error in a docstring/assertion in the Mamba attention backend and a style violation in the FlashInfer backend. Overall, the changes look good and address the intended purpose.

[gemini-code-assist]

The backslash \ for line continuation is redundant here because the expression is already inside parentheses. According to PEP 8, implied line continuation within parentheses is preferred. Removing the backslash will make the code cleaner and more compliant with standard Python style.

Suggested change

	use_cudagraph = (self.enable_cuda_graph and pure_decode and \
	num_decodes <= self._decode_cudagraph_max_bs)
	use_cudagraph = (self.enable_cuda_graph and pure_decode and
	num_decodes <= self._decode_cudagraph_max_bs)

[gemini-code-assist]

The docstring and assertion message here refer to 'MLA' (Mixture-of-LoRA-Experts Attention), but this file is for Mamba attention. This seems to be a copy-paste error and could be confusing for future developers. Please update them to refer to 'Mamba'.

        """
        This method builds the metadata for full cudagraph capture.
        Currently, only decode is supported for full cudagraphs with Mamba.
        """
        m = common_attn_metadata
        assert m.num_reqs == m.num_actual_tokens, \
            "Mamba only supports decode-only full CUDAGraph capture. " \
            "Make sure all cudagraph capture sizes <= max_num_seq."