[Feature] Support paged flash attention for inference kv cache

[npuichigo]

I'm working on a burn-based vllm but it's slightly harder to support this part. Do we have any plan to support paged attention with block_table support in burn directly?

[laggui]

It's not currently on the roadmap. Flash attention is not entirely supported yet, this extension would also have to be added via cubek-attention.