[Bugfix] Bandaid fix for speculative decoding tests

[tlrmchlsmth]

In #6484, the attention selection was changed to allow for attention-free models like Mamba to have a placeholder attention backend, as Mamba still needs attention metadata to manage its internal state.

The change was to call get_attn_backend even if num_attn_heads was zero. This is causing a divide-by-zero during self.model_config.get_head_size(). This PR skips get_attn_backend when num_attn_heads is zero, unless self.model_config.is_attention_free is set, which is OK is_attention_free is only True for Mamba at the moment.

This is a bandaid fix to get the build green, as this is a little hacky.

[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 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 do one of these:

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

🚀