fix: call fa3 to do mla instead of flashmla

[tianhaox]

Overview:

for regular path deepseek, on hopper, it by default calls fa3 to do mla for prefill and decode.
On Blackwell, it calls trtllm-gen kernels. But this is not verified yet. All code is verified on hopper.

Limitations:

1. the efficiency of fa3 for context mla is not that good for GQA, 30-50% perf of GQA.
2. generation attention fp8 kvcache perf is very poor. https://docs.sglang.io/advanced_features/attention_backend.html
3. current mla kvcache pool op has bug. It uses int32 for indexing and it limits bs*seq_len up to ~4M tokens. Modify the collector a bit to remove those test cases.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[davilu-nvidia]

LGTM