tokenspeed-attention: add TokenSpeed MHA/MLA Triton kernels

[adarshxs]

What does this PR do?

Adds tokenspeed MHA/MLA triton kernels

Summary

Adds a new tokenspeed-attention Hub kernel package with portable Triton attention kernels ported from LightSeek TokenSpeed.

Exported APIs:

• mha_prefill
• mha_extend_with_kvcache
• mha_decode_with_kvcache
• mla_prefill
• mla_decode_with_kvcache
• attn_merge_state

[danieldk]

/kernel-bot build-and-stage tokenspeed-attention

[github-actions]

Build request processed.

Command: /kernel-bot build-and-stage tokenspeed-attention
Mode: build and stage
Target branch: pr-990
PR head SHA: 2f6e9a9b6e5560a8c4de3626bedda75127b58f95
Workflows: build.yaml, build-mac.yaml, build-windows.yaml

Dispatched (2):

• tokenspeed-attention (build-mac.yaml): https://github.com/huggingface/kernels-community/actions/runs/28027306580
• tokenspeed-attention (build.yaml): https://github.com/huggingface/kernels-community/actions/runs/28027307719