[2/8] refactor(kernels): move attention kernels to miles/kernels/attention

[Zhichenzzz]

What

Files the attention kernels under miles/kernels/attention/. Moves and import updates only.

• dsa/: the GLM-5 / DeepSeek-V3.2, DeepSeek-V4 and DeepSeek-V4.1 TileLang indexer + sparse-attention copies (merged in [3/8] refactor(kernels): unify the DSA kernels of GLM-5 and DeepSeek-V4 / V4.1 #3607), the top-k selector, and GLM-5.3-Flash's key-pool selection (kpool.py).
• qsa/: Qwen3.8-Next's Triton block-sparse attention.
• README.md: the layout and the rules.

The GDN backend selector stays in miles_plugins/models: it only picks between fla and FlashQLA and owns no kernel. DeepSeek-V4.1's dsa_topk import follows the selector; left alone, it would break at import time without any textual conflict.

Tests

• Pre-rebase run on H200: the GDN backend test, both manual DeepSeek-V4 TileLang suites (indexer and sparse MLA) and the GLM-5 / DSv4 model imports give results identical to main.
• pre-commit clean.

This branch on H200 (45f9d98bc), against main

Test	This branch	main
py_compile + import of every changed module	pass	–
tests/fast/test_linear_attn_kernel_selection.py	3 passed	3 passed
DeepSeek-V4 TileLang indexer suite	45 passed	45 passed
DeepSeek-V4 TileLang sparse MLA suite	23 passed	23 passed

Stack

	PR	Change
1	#3605	non-attention kernels → miles/kernels
2	#3606	attention kernels → miles/kernels/attention
3	#3607	one DSA kernel pair (GLM-5, DeepSeek-V4, V4.1)
4	#3608	per-arch DSA sparse attention dispatch
5	#3609	shared head-sharded linear-attention layer (miles_plugins/models/linear_attn.py)
6	#3610	Qwen3.5 / Qwen3-Next GDN on it
7	#3632	Kimi-K3 KDA on it
8	#3634	short-conv and KDA backward backends

Rules for miles/kernels are in its README: only kernels we write, one directory per op, names say what is computed (never the backend), callers import the entry, a torch-reference test per op.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[HJSang]

LGTM. Just refactoring and no risk to merge