[WIP] Llama 4: Hybrid KV buffer (disable radix attention)

[tarinkk]

Motivation

We set KV buffer of different sizes for global and local attention layers in Llama 4 for better memory usage.

Modifications

We modify the size of KV buffer in MHATokenToKVPool and set two TokenToKVPoolAllocator for global and local attention.
The hybrid ratio between 0 to 1 can be set by --enable-hybrid-kvcache. default set is 1.0 (0 = pure uniform: local_size / global_size = 1, 1.0 = pure hybrid: local_size / global_size = local_attention_size / context_length)
Currently, only support cases when page size = 1, disable radix attention and disable cuda graph.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.