[torch.compile] add dynamo time tracking

[youkaichao]

$ vllm serve meta-llama/Meta-Llama-3-8B
Graph capturing finished in 10 secs, took 0.32 GiB
init engine (profile, create kv cache, warmup model) took 14.18 seconds

$ vllm serve meta-llama/Meta-Llama-3-8B -O3
Dynamo bytecode transform time: 4.60 s
Compiling a graph for general shape takes 14.77 s
torch.compile takes 19.37 s in total
Graph capturing finished in 15 secs, took 0.33 GiB
init engine (profile, create kv cache, warmup model) took 39.34 seconds

This PR adds the missing piece of the compilation time.

By adding -O 3, the init time increases by 39.34 - 14.18 = 25.16 seconds, and the breakdown analysis is:

• Dynamo bytecode transform time: 4.60 s
• Compiling a graph for general shape takes 14.77 s
• triton compilation (shown by the increase of cudagraph capture time, which includes warmup triton kernels) takes 5 seconds

[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.

🚀