[Torch] Add aten.diag decomposition

[alex1xu]

torch.diag(A) fails to compile through the torchdynamo-export-to-torch-backend pipeline with error: failed to legalize operation 'torch.operator' that was explicitly marked illegal (KernelBench level1/12, return torch.diag(A) @ B). The FX importer has no first-class op for aten.diag, so it emits a generic torch.operator "aten.diag"; LowerToBackendContract marks all such ops illegal and aborts because nothing resolves or decomposes it.

This adds aten.diag as a first-class op and decomposes it. torch.diag is rank-dependent (PyTorch aten/src/ATen/native/TensorShape.cpp): a 1-D input builds a 2-D matrix with the input on the diagonal-th diagonal, and a 2-D input extracts the diagonal-th diagonal as 1-D. DecomposeAtenDiagOp mirrors that dispatch: rank 1 lowers to aten.diag_embed(self, offset=diagonal, dim1=0, dim2=1), rank 2 to aten.diagonal(self, offset=diagonal, dim1=0, dim2=1), with an early notifyMatchFailure for any other rank. Reusing the existing diag_embed/diagonal lowerings rather than adding a new TorchToLinalg pattern keeps the change to a decomposition and inherits those ops' backend coverage.

• ODS: aten::diag in torch_ods_gen.py; shape/dtype in abstract_interp_lib_gen.py (aten〇diag〡shape/〡dtype), regenerated into the checked-in GeneratedTorchOps.td / AbstractInterpLibrary.cpp.
• Registration: addPatternIfTargetOpIsIllegal<DecomposeAtenDiagOp> and target.addIllegalOp<AtenDiagOp>() in LowerToBackendContract.cpp.

aten.diag with a nonzero offset on a 1-D input routes to aten.diag_embed, whose existing linalg lowering does an out-of-bounds extract for nonzero offset (the same crash that already lists AtenDiagEmbedOffsetDiag in LINALG_CRASHING_SET); Diag1DOffsetModule is added there for parity, and fixing that lowering is left as separate work.

Testing

• test/Dialect/Torch/decompose-complex-ops.mlir: diag_1d (offset 2, lowers to diag_embed), diag_2d (offset 3, lowers to diagonal), and diag_rank3_no_decompose (rank-3 left intact). Offsets are chosen distinct from the dim1=0/dim2=1 constants so FileCheck pins operand wiring rather than matching coincidentally.
• aten〇diag〡shape carries a @check_shape_function that validates the 1-D/2-D, positive/negative-offset, empty-result, and rank-3-error cases against eager torch.diag at library-generation time.
• e2e (projects/pt1/python/torch_mlir_e2e_test/test_suite/diagonal.py): Diag1DModule, Diag1DOffsetModule, Diag2DModule, Diag2DNegativeOffsetModule, with xfail_sets.py updated — ONNX has no aten.diag import path; stablehlo mirrors the existing diag_embed/diagonal xfails; the 1-D nonzero-offset case is LINALG_CRASHING.

Part of #4575.