Require Support For DeepSeek-MLA and batchsize > 1

FlexPrefill is an amazing work for longtxt tasks

Thanks for the authors, we've make it into sglang and save 32% time for RULER and NIAH benchmarks, Qwen3-235B-FP8 on 4 H20.

But now we have 3 problems, looking for help

* Support DeepSeek-MLA
We've adjust the dim to meet the MLA's [B * N, H, 192], [B * N, H, 192], [B * N, H, 128] in sglang's `forward_extend`, but the RULER accuracy seems drop so hard, (the niah_multikey_1 drops from 1.0 to 0.2 ... )

* Support batchsize > 1
In longtxt scenes, usually the batch_size is not too large, but usually goes to 2 or 4, so the batch_size > 1 is requiring support.

* Compatibality with sglang's --chunked-prefix-cache feature
Another thing is about the sglang's --chunked-prefix-cache feature, it requires Attention kernel to return a `lse` to support cache for long sequence, while FlexPrefill kernel does NOT return lse. Does FlexPrefill can work with this feature ?

Can you support for these 3 problems ?