fla: sync with flash-linear-attention main

[kashif]

Important: Read before submitting

New contributors: Please open or comment on an issue before submitting a
PR to discuss the change you'd like to make. This helps us align on approach and
avoids wasted effort on changes we may not be able to merge. Please only create
a PR once one of the project maintainers agrees on your outlined approach.

PRs of contributors who are not vouched for are automatically closed. Regular
contributors are added to the vouch list.

The kernels-community repository is for:

1. Kernels developed by Hugging Face and partners.
2. Kernels developed by third parties, that have not been 'kernelized' yet, but
   are used by Hugging Face projects such as diffusers and transformers.

We cannot accept PRs for the following:

• New kernels that are not in categories (1) and (2).
• Changes to kernels in category (2), these should be submitted upstream.

For, LLM-generated changes, we prefer that you write your prompt in an issue
over a PR with LLM-generated changes.

Related issue

Closes #

What does this PR do?

Syncs the vendored fla kernel with the current flash-linear-attention main (e52dbc0).

Motivation

fla was a couple months stale. Routine refresh.

Changes

• Updated fla/torch-ext/fla/{ops,modules,utils} from upstream, new absolute imports rewritten to relative as usual.
• New op families came along: atk, momentum_delta_rule, precond_gated_delta_rule, precond_kda, and a few more.
• deltaformer needs pad_input/unpad_input from fla/layers/utils.py upstream (not vendored here) — originally pulled just those two functions into fla/torch-ext/fla/utils/padding.py; rebasing onto main picked up fla: fix relative import depths in the vendored source and add smoke tests #1129's fla/torch-ext/fla/utils/layers.py (a full vendored copy, same idea, landed independently while this was open), so dropped padding.py and pointed deltaformer at that instead.
• __init__.py now sets __version__ (upstream's _decorators.py needs it), kept everything else about it and layers.py as-is.
• Bumped build.toml version 1 → 2 (on top of main's xpu backend addition), check_public_api.py flagged new optional kwargs on chunk_gated_delta_rule/chunk_kda.
• Dropped triton_ascend backends everywhere (NPU-only, build.toml declares backends = ["cuda", "xpu"], no NPU) and tilelang backends that were pure speed optimizations (rwkv6, dplr, kda, and now gdn2 too — upstream added a triton_ascend gdn2 backend in the same window). Kept ops/common/backends/tilelang (+ its attn/common registrations) since it's the only fallback for a real Hopper correctness bug on Triton 3.4.0-3.7.0, not just perf (see the RuntimeError in chunk_o.py, upstream metal-flash-sdpa: add or update card #640).
• Also swept the remaining scattered NPU-only code: ascend_compile_kwargs() and a triton extra.cann→extra.ascend alias in utils/_compat.py, and an Ascend grid-size workaround in modules/token_shift.py. Left the plain IS_NPU/IS_AMD/IS_INTEL device-platform flags alone — shared, cheap, zero-dependency, not NPU-specific code.
• Rebased onto main to pick up fla: fix relative import depths in the vendored source and add smoke tests #1129 (relative-import fixes + fla/tests/test_fla.py, both kept as-is) and fla: add XPU support and optimizations #1145 (XPU support — hand-merged its IS_INTEL additions in chunk_o.py/chunk_delta_h.py/gated_delta_rule/{chunk_fwd,wy_fast}.py together with this branch's own upstream sync of those same files).
• Picked up the small amount of upstream flash-linear-attention movement since the last sync (516143e→e52dbc0): gdn2 gained @dispatch decorators (its new triton_ascend backend isn't vendored, same as everywhere else — dispatch() no-ops cleanly with no registry present), and solve_tril.py/utils/_device.py/utils/__init__.py picked up real AMD gfx1250 TMA support.

Testing

• Imported every submodule of the built package. Everything imports cleanly except two pre-existing issues, neither introduced here (reproduced identically against flash-linear-attention itself): a missing/mismatched tilelang install, and a real upstream bug in ops/delta_rule/parallel.py (imports a name wy_fast.py doesn't define).
• Re-ran check_public_api.py after every change — same two additive kwarg changes throughout, nothing new.

Checklist

• This PR is linked to an issue that was discussed and approved
• I have tested these changes locally
• New/changed functionality has test coverage
• LLM disclosure:
  • I did not use an LLM to create this PR.
  • I used and LLM for assistance while creating this PR.
  • This PR was mostly or completely generated by an LLM.

🤖 Generated with Claude Code

[vasqu]

Waitng on #1129 first so we can make sure we have tests for gdn and kda

I guess this sync would also be mainly for kda and gdn? I recall some support for context parallel, can we check in layers.py as well to sync the compatibility more concretely