[V1][TPU] Do not compile sampling more than needed

[NickLucche]

We're compiling a few extra graphs for sampling that we should never come across at runtime.
Namely the number of requests that require sampling can never exceed the number of tokens that are generated.
Eg. when we get N=16 tokens in input, we are not going to get 128 sequences for sampling. So only compile up to N.

Assuming max number of sequences = 32

INFO 04-01 11:04:44 [tpu_model_runner.py:839] Compiling sampling with different input shapes.
INFO 04-01 11:04:44 [tpu_model_runner.py:860]   -- num_tokens: 16, num_seqs: 8
INFO 04-01 11:04:44 [tpu_model_runner.py:860]   -- num_tokens: 16, num_seqs: 16<==
INFO 04-01 11:04:46 [tpu_model_runner.py:860]   -- num_tokens: 32, num_seqs: 8
INFO 04-01 11:04:47 [tpu_model_runner.py:860]   -- num_tokens: 32, num_seqs: 16
INFO 04-01 11:04:48 [tpu_model_runner.py:860]   -- num_tokens: 32, num_seqs: 32

[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.

🚀

[yaochengji]

LGTM, thanks!