[V1] Add KV cache group dimension to block table

[heheda12345]

This PR adds the KV cache group (KVCacheConfig.groups introduced by #11960) dimension to block table, to prepare for supporting allocating different blocks to different layers.

It is splitted from #11938 and is a preparation for #11382

Benchmark result ($\Delta$ is compared with main branch df450aa, shows the minimum time of 3 runs, H100):
commands:

VLLM_USE_V1=1 python3 -m vllm.entrypoints.openai.api_server --port 8888 --disable-log-requests [--model facebook/opt-125m, --model meta-llama/Llama-3.2-1B-Instruct, --model meta-llama/Llama-3.1-8B-Instruct]
python $VLLM_DIR/benchmarks/benchmark_serving.py --port 8888 --ignore-eos --seed 555 --dataset-name sharegpt --dataset-path /data/zhang-chen/dataset/ShareGPT_V3_unfiltered_cleaned_split.json [--request-rate 70 --model facebook/opt-125m, --request-rate 70 --model meta-llama/Llama-3.2-1B-Instruct, --request-rate 30 --model meta-llama/Llama-3.1-8B-Instruct] --percentile_metrics='ttft,tpot,itl,e2el'

Model             E2EL(ms)   ∆E2EL(%)    ITL(ms)    ∆ITL(%)   TPOT(ms)   ∆TPOT(%)   TTFT(ms)   ∆TTFT(%)
opt-125m            437.16     +1.10%       1.87     +1.08%       1.90     +1.60%      14.92     -0.13%
llama-1b            796.35     +0.61%       3.91     +0.51%       4.00     -1.23%      24.43     -0.41%
llama-8b           2652.95     +0.00%      13.23     +0.00%      13.36     +0.00%      40.08     +1.34%

It shows that adding the group dimension does not introduce much overhead.

[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.

🚀

[heheda12345]

get block_size from the KVCacheSpec of each group instead.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @heheda12345.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork