Port the PaddlePaddle deltas from paddle/v1.27.0 onto 1.28.0 - #12

Carry the PaddlePaddle deltas from paddle/v1.27.0 over to the new
paddle/v1.28.0 branch. Without them Paddle cannot drive these kernels at all
(the dtype tables resolve against PyTorch, TensorDesc rejects a Paddle place,
torch.memory_format does not exist, numel() becomes a blocking D2H, ...).

Base is paddle/v1.28.0 = upstream 1.28.0-rc @ 601b3f35f, which has the same
tree (b4b9b06cc) as the v1.28.0 release tag, so the branch is pristine
upstream 1.28.0 and this PR is exactly the Paddle delta on top of it.

Produced with git rebase --onto <1.28.0 base> 0018f8dff paddle/v1.27.0, so every
original commit keeps its author, message and granularity.

Commits

Two of the nine paddle/v1.27.0 commits are not here, because on this base
they carry no content:

• 9d82b45b (Compat with PaddlePaddle) only added a Paddle-local nvtx_range
  context manager to _interface_sm100.py and indexer_backward_sm100.py, which
  add compat for nvtx.range and ExternalStream #2 then replaced with the generic torch.cuda.nvtx.range shim in
  utils/runtime.py. 1.28.0 already calls torch.cuda.nvtx.range, so the commit
  rebases to empty and git dropped it. The fork's net change to
  _interface_sm100.py is likewise zero.
• Optimize DSA backward SM100 with enhanced topk_length loop and TMA fast path #5 eb129d41 (DSA backward topk_length loop + TMA fast path) was reverted by
  Fix v1.27.0 error #8 2aef928b, so it has no net effect either. Replaying it would have meant
  hand-porting 315 lines onto the kernel upstream rewrote in Improve performance of SM100 DSA backward kernel NVIDIA/cudnn-frontend#684 only to revert
  them again — and Fix v1.27.0 error #8's revert restores the 1.27.0 file, which would clobber
  that rewrite. Skipped, which leaves Fix v1.27.0 error #8 as just its
  indexer_backward/api.py line (view(1) → reshape([1])).

The v1.27.0 indexer_backward_sm100.py performance regression (range_constexpr
→ range(unroll_full=True), upstream NVIDIA#376) is already fixed in 1.28.0 by
upstream NVIDIA#549, so nothing is carried for it.

Conflict resolutions worth reviewing

Upstream reworked the code four of the fork hunks sat on:

1. datatypes.py _torch_to_cutlass_data_type — 1.28.0 resolves the
   framework via sys.modules.get("torch") instead of importing it. A real
   PyTorch is commonly installed alongside Paddle, so that probe hands back
   PyTorch and every Paddle dtype misses the mapping. Changed to probe
   "paddle", which keeps upstream's "do not import on behalf of another
   framework" property. This one is load-bearing: without it
   _torch_to_cutlass_data_type(paddle.bfloat16) returns None and every
   downstream dtype check silently degrades.
2. api_base.py TensorDesc.__post_init__ — gained the same _torch()
   probe plus a framework-neutral Device. Probes Paddle directly and keeps
   Device in the accepted set.
3. api_base.py is_contiguous — the v1.27.0 workaround is resolved
   in favour of upstream. The old signature was memory_format: torch.memory_format = torch.contiguous_format, which fails at def time
   because Paddle has no memory_format; upstream changed it to Any = None
   and guards every torch reference, so the raise NotImplementedError stopgap
   is no longer needed and the contiguity check works again under Paddle.
4. dsa_bwd_sm100.py — rewritten upstream by Improve performance of SM100 DSA backward kernel NVIDIA/cudnn-frontend#684, so [Fix] Protect prob <=1 when cudnn backward #4's clamp landed at
   the new exp2 site by hand rather than as a context-matched patch.

Commit 8 is the one change that is not a port: dsa_bwd_sm100_h16.py is new
in 1.28.0 (NVIDIA#664, added after v1.27.0 was cut), is a copy of the same epilogue and
carries the identical unclamped exp2. Drop that commit if you would rather keep
the branch a strict port.

Commit 9 exists because the ported Paddle code predates the current
.pre-commit-config.yaml (black, --line-length 160) and six files were not
format-clean; the 1.28.0 base already was.

Verification

black --line-length 160 --fast clean over all 18 touched files. No C/C++/CUDA
files touched, so clang-format is a no-op. All 18 files byte-compile.

Smoke-tested on B30Z (sm_103) with Paddle 3.4.0 under
paddle.enable_compat(scope={"cudnn"}), which is how paddlefleet_ops drives
this package:

• import cudnn succeeds and the eager api_base / DSA imports resolve (23 DSA symbols)
• _torch_to_cutlass_data_type: paddle.bfloat16 → BFloat16, paddle.float32 → Float32
• _is_torch_tensor(paddle_tensor) → True; to_cute_tensor builds a cute tensor
• _validate_grad_loss_tensor accepts (), (1,), (1,1) → shape (1,); rejects numel 2 and fp16; and with Tensor.numel monkeypatched to raise, still passes — i.e. the guard is provably host-side
• TensorDesc accepts a Paddle place and is_contiguous() returns True
• resolve_stream / torch_stream_context round-trip

Not verified: no kernel was actually launched, so the #9 dK barrier and the #4
clamp are carried on code review plus their original v1.27.0 validation, not on a
fresh numerical run.

Merge requirements

merge-requirements currently fails for two reasons, and only one of them is
mine to fix:

• this PR has no Milestone — please set it from the sidebar;
• the repository secret PROJECT_READ_TOKEN is not configured, so the workflow
  cannot read the Projects field and fails closed. That affects every PR on this
  repo, not just this one, and needs a repo admin (or the
  cat-routine-update label as an exemption).