Force cudaGraphExec reinstantiation when clusters are used

[andportnoy]

Proposed changes

Thread block cluster dimensions are not correctly updated by cudaGraphExecUpdate. Therefore, when clusters are used, we reinstantiate a cudaGraphExec rather than updating it.

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
  The change fixes existing tests.
• I have updated the necessary documentation (if needed)

[awni]

Some benchmarks on B200

Inference with meta-llama/Meta-Llama-3.1-8B

Pre: prompt_tps=59185.146
Post: prompt_tps=59359.110

Pre: generation_tps=282.392
Post: generation_tps=282.671

Pre: toks_per_sec: 97961.0518
Post: toks_per_sec: 97452.2300

[awni]

@andportnoy I changed this a bit more than I expected. Basically three changes:

1. we don't bother cache the graph exec if the graph is not updatable since there is no point
2. The condition for allowing the update are less conservative (if there is a non-singleton cluster in the x dimension of a sub graph with a single kernel node we can still update the graph).
3. In order to achieve the above I encoded the cluster x dimension in the graph key which took a bit of rearranging to make it work nicely.

[awni]

This resolves outstanding test failures! Thanks for the fix!!