[V1] Fix Compilation config & Enable CUDA graph by default

[WoosukKwon]

This PR fixes a performance bug on V1 introduced by #10437, which disabled custom ops even when torch.compile was not used.

Also, this PR enables the piecewise CUDA graphs by default since #10237 was merged.

[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.

🚀

[ProExpertProg]

Is there a reason why this doesn't just use total_gpu_memory from after the profile run (like it was done before)?

[ProExpertProg]

I think this is too pessimistic, if anything else is allocated on the GPU (like when we want to do 2 llm instances in tests), this will count all that memory as if it was allocated in the forward pass. I think we should instead just subtract 2 values here: #18974

[laithsakka]

This PR override the compilation mode that the user provides is that expected?
Shall we warn at least ?
for example
If i am doing something like

conf = vllm.config.CompilationConfig(level=CompilationLevel.NO_COMPILATION )
llm = LLM(model="meta-llama/Llama-3.2-1B",compilation_config=conf)

then level will be overriden by this code when VLLM_USE_V1 is on. and go back to become PIECEWISE

  if envs.VLLM_USE_V1 and self.model_config is not None and \
            not self.model_config.enforce_eager:
            # By default, V1 uses piecewise CUDA graphs. If full_cuda_graph
            # is set to True, full CUDA graphs will be used.
            self.compilation_config.cudagraph_num_of_warmups = 1
            self.compilation_config.level = CompilationLevel.PIECEWISE
            self.compilation_config.set_splitting_ops_for_v1()

[ProExpertProg]

Yes, I believe #19340 is supposed to address this.