[Qiming Bao (14H034160212)]

aten_cdist is still a NotImplementedError stub, so torch.onnx.export(..., dynamo=True) fails on any model that calls torch.cdist:

DispatchError: No ONNX function found for <OpOverload(op='aten._cdist_forward', overload='default')>

(#1647 was closed, but cdist was not implemented.)

This PR implements it for aten::cdist and aten::_cdist_forward, following PyTorch's semantics:

• p == 2: uses the matmul form sqrt(max(|x1|² + |x2|²ᵀ − 2·x1·x2ᵀ, 0)), as PyTorch does, for compute_mode=1, and for the default mode when P or R is greater than 25 or not known statically. This avoids materializing the [*, P, R, M] difference tensor for large inputs. Otherwise, and for compute_mode=2, it uses the exact difference form.
• Other p: reduces |x1[..., :, None, :] − x2[..., None, :, :]| over the last axis with ReduceL1 (p=1), ReduceL2, ReduceMax/ReduceMin (±inf), a non-zero count (p=0) or Pow/ReduceSum.
• Batch dimensions broadcast as in PyTorch.

Tests:

• Added cdist to ops_test_data.py. All 120 non-empty OpInfo subtests pass (p ∈ {1, 1.5, 2, 2.5, 3, inf}, both compute modes, broadcasting batch dims). Size-0 samples are skipped with a fixme, like flip and pixel_unshuffle: ORT returns the wrong shape when reducing a size-0 input.
• End-to-end through torch.onnx.export(dynamo=True) + ONNX Runtime: 146/146 cases match eager (p ∈ {0, 1, 1.5, 2, 3, inf} × three compute modes × small/large/broadcast shapes × float32/float64, plus dynamic P and R).
• ruff check / ruff format (0.15.8) clean.