flash-attn-hopper: bump to 00756db for TE 2.17 window kwargs

[yueming-yuan]

Problem

miles bumped transformer_engine to 2.17.0 (radixark/miles#1781). TE 2.17 moved
use_flash_attn_3 out of the window_size-tuple branch in
cp_p2p_fwd_flash_attn / cp_p2p_bwd_flash_attn:

# TE 2.12
if use_flash_attn_3 or (v2_3_plus and not v2_7_0_plus):
    fa_forward_kwargs["window_size"] = (-1, -1)        # FA3 took the tuple
elif v2_7_0_plus:
    fa_forward_kwargs["window_size_left"] = -1

# TE 2.17
if v2_3_plus and not v2_7_0_plus:
    fa_forward_kwargs["window_size"] = (-1, -1)
elif use_flash_attn_3 or v2_7_0_plus:                  # FA3 moved here
    fa_forward_kwargs["window_size_left"] = -1

The pinned revision fbf24f67 builds flash_attn_3 3.0.0b1, whose
_flash_attn_forward only accepts a single window_size tuple, so every
context-parallel path raises TypeError. There is no FA3-only kill switch in
TE — NVTE_FLASH_ATTN disables FA2 and FA3 together.

FA2 in the same image (2.7.4.post1) already takes the new kwargs, so only the
FA3 path is affected.

Change

• Bump the hopper/ pin to 00756db, which builds 3.0.0.
• Force a source build. Without it setup.py prints
  Guessing wheel URL: .../v3.0.0/flash_attn_3-3.0.0+cu122torch2.11...whl and
  downloads a prebuilt wheel built against a different CUDA/torch than the image.
• Drop the separate flash_attn_interface.py fetch in test_wheels.py. It was a
  second pin that can drift from the one the extension module was built at. The
  wheel now ships the interface top-level and a flash_attn_3.flash_attn_interface
  re-export shim, so nothing extra is needed — and copying the full module over
  that shim would re-run @torch.library.custom_op("flash_attn_3::...") and
  double-register the ops.

Verification

Built and installed on an h200-sci devbox running the radixark/miles:pr-1795
image itself (torch 2.11.0+cu130, CUDA 13.0), so the wheel is compiled against
the toolchain it will ship with.

[INFO     | DotProductAttention]: Running with FlashAttention backend (version 3.0.0)
[PASS] all 5 symbols TE imports exist
[PASS] _flash_attn_forward schema takes window_size_left/right  ['SymInt window_size_left=-1', 'SymInt window_size_right=-1']
[PASS] _flash_attn_backward schema takes window_size_left/right  ['SymInt window_size_left=-1', 'SymInt window_size_right=-1']
[PASS] import transformer_engine.pytorch
[PASS] TE sees flash-attn-3 == 3.0.0  fa3_version=3.0.0 v3_0_0_beta=False
[PASS] TE bound a v3 forward
[PASS] TE's bound v3 fwd is the same op we checked
[PASS] fwd accepts window_size_left/right kwargs  out=(2, 512, 8, 128) lse=(2, 8, 512)
[PASS] returns (out, softmax_lse, ...) in the order TE indexes [0]/[1]
[PASS] TE DotProductAttention fwd+bwd on FlashAttention backend  out=(2, 512, 1024) grad_finite=True
[PASS] output matches SDPA reference  max_abs_err=0.001953
[PASS] FA3 registered as installed in TE
       fa2=2.7.4.post1 fa3=3.0.0 v3_installed=True

RESULT: ALL PASS

Notes on things checked so they are not rediscovered later:

• The five symbols are load-bearing. TE guards only PackageNotFoundError
  around the version lookup, not ImportError around the imports themselves, so
  a wheel missing any of flash_attn_func, flash_attn_varlen_func,
  flash_attn_with_kvcache, _flash_attn_forward, _flash_attn_backward makes
  import transformer_engine.pytorch raise rather than degrade to cuDNN.
• 3.0.0b1 -> 3.0.0 is benign. It flips
  FlashAttentionUtils.v3_0_0_beta to False; the only consumer
  (backends.py:1230) uses it to append an "update your flash-attn v3 beta"
  hint to an already-raised TypeError. No functional branch depends on it.
• Arch coverage is unchanged: sm_80 + sm_90a, verified with cuobjdump
  against both the old and new _C.abi3.so. hopper/setup.py derives arch from
  source-file suffixes (_sm90.cu -> sm_90a) and ignores the
  TORCH_CUDA_ARCH_LIST that _setup_env sets, so that variable has never had
  any effect on FA3. FA3 has no Blackwell kernels here either way.
• Return shape is compatible. 3.0.0 returns
  (out, softmax_lse, out_accum, softmax_lse_accum); TE's CP code reads indices
  [0] and [1] under use_flash_attn_3, which line up.