candle-flash-attn: update vendored Tri Dao FA kernels to v2.8.3

[toddwbucy]

Closes the kernel-vendoring portion of #3515.

Summary

Bumps the vendored Flash Attention 2 kernel sources from the
post-Dec-2024 state (~v2.6.x with #2688 / #2689 / #2690 cherry-picks)
to upstream v2.8.3 (Tri Dao commit 060c918, 2025-08-14).

Single squashed commit. ~700 net lines including 24 new split-KV
kernel files; the bulk is mechanical kernel updates from upstream
v2.8.3.

What's in the bump

• All kernels/*.h and forward sm80 *.cu files replaced with
  v2.8.3 sources, path-remapped from upstream
  csrc/flash_attn/src/ to candle's flat kernels/.
• 24 new split-KV forward sm80 kernels added (flash_fwd_split_hdim*_*_sm80.cu)
  — {fp16, bf16} × {dense, causal} × {hdim 32, 64, 96, 128, 192, 256}.
• flash_api.cu dispatcher mirrors v2.8.3's
  run_mha_fwd(params, stream, force_split_kernel) shape: branches
  between run_mha_fwd_<> (dense) and run_mha_fwd_splitkv_dispatch<>
  based on params.num_splits and force_split_kernel.
• extern "C" run_mha FFI gains four params (num_splits,
  softmax_lseaccum_ptr, oaccum_ptr, force_split_kernel).
  src/ffi.rs mirrors. Existing call sites pass safe defaults so
  pre-bump behavior is preserved exactly.
• Rust-side ports of upstream's num_splits_heuristic (efficiency
  search, ≥85% of peak occupancy, capped at 128) and the
  set_params_splitkv accumulator-allocation logic.
• Dropped legacy hdim 160 / 224 / 512 forward kernels — Tri Dao
  removed their launch-template helpers in v2.x and the previously
  vendored .cu files for those head dims no longer compile against
  v2.8.3's launch template.
• CUTLASS pin retained at 7d49e6c7 (v2.8.3's pinned dc481792
  is not required for the kernels we compile).

PyTorch dependency stub

v2.8.3's dropout codepath references at::PhiloxCudaState and
at::cuda::philox::unpack. Provide a candle-side stub in
kernels/philox_unpack.cuh so the path compiles without dragging
the torch C++ runtime in. Tripwire at the dispatch site asserts
params.p_dropout == 1.f to prevent any future caller from
silently triggering the (broken-by-stub) dropout codepath.

Compile-time fix: extern template declarations

The naive splitkv-aware flash_api.cu produced a runaway nvcc
compile (~30+ minutes, ~17 GB cicc RSS) during development. Root
cause: flash_fwd_launch_template.h defines the primary template
run_mha_fwd_splitkv_dispatch<>, whose body has 7 nested
*_SWITCH chains (128 kernel specialisations per
(dtype, hdim, causal) tuple, plus a 14-way combine-kernel chain).
Implicitly instantiating all 24 tuples in the dispatcher TU
expanded ~3,400 kernel specialisations through CUTLASS GEMM
templates in a single nvcc invocation.

Fixed by adding 24 extern template declarations in flash_api.cu
after the include. nvcc no longer instantiates the splitkv
specialisations in this TU; linker resolves to the explicit
instantiations in the per-hdim flash_fwd_split_hdim*_*_sm80.cu
files (which compile in parallel as 24 ~30s TUs). The dense-path
counterpart run_mha_fwd_<> was already forward-declared in
flash.h without a primary template definition, so it never had
this problem.

Testing

cuda:0, sm_86, RTX A6000:

$ cargo build           # 4m36s clean (49 kernels)
$ cargo test
  flash_attn_acausal          ok
  flash_attn_acausal_softcap  ok
  flash_attn_varlen           ok
  flash_attn_acausal_splitkv  ok   # NEW

The new flash_attn_acausal_splitkv test uses a shape
(batch=1, heads=2, seqlen_q=8, seqlen_k=512, head_dim=64) chosen
so num_splits_heuristic returns ≥ 2 on any modern sm80+ GPU,
and asserts max-abs < 5e-3 against an fp32 attention reference.

Public Rust API

flash_attn, flash_attn_windowed, flash_attn_alibi,
flash_attn_varlen, etc. — unchanged. New FFI params are internal.

Note on framing

Originally filed under the hypothesis "FA kernels are stale,
bumping fixes long-context divergence." The bump is structurally
fine and worth landing on its own merits (newer kernels + splitkv
support + no regressions), but it is correctness-neutral on
long-context by itself. The actual long-context fix turned out
to be in candle-transformers/src/models/qwen2.rs::RotaryEmbedding::new
(filed as a separate PR, link to follow if maintainers want to
review them together). See the diagnostic update on #3515
for the full chain.

This PR is the "newer kernels + splitkv" half of the original ask.

[toddwbucy]

Cross-link: #3520 is the actual long-context correctness fix (5-line dtype change in qwen2.rs RoPE). Originally framed in #3515 as one issue; the diagnostic chain split them into two independent contributions. Either review order works.