[Frontend] Add Comprehensive Core Aten IR Operator Support

[sdh1014]

Background

PyTorch exposes more than 2000 operators. Implementing a custom lowering for every single operator in a backend would be prohibitively expensive. Fortunately, PyTorch provides decomposition facilities: complex high-level operators in PyTorch (such as torch.nn.functional.softmax or torch.batch_norm) can be decomposed and normalized by the AOTAutograd component into Core Aten IR.

According to the official documentation for torch.compiler_ir (https://docs.pytorch.org/docs/2.8/torch.compiler_ir.html), decomposed graphs are expressed using both Core Aten IR and Prims IR. Compared to Core Aten IR, Prims IR is a lower-level, more fine-grained IR.

On PyTorch 2.8.0, this change targets all 191 Core Aten IR operators listed in that document and provides full coverage for them in our backend.

---

New Operator Implementations

Frontend Core (frontend.py, operation.py)

• Added new Op classes for Core Aten IR operators.
• Updated operator mappings for renamed ops (e.g., LessThanOp → LtTensorOp, GreaterThanOp → GtTensorOp).
• Added mapping for the _embedding_bag_forward_only operator.
• Updated documentation for the aot_autograd_decomposition parameter.

TOSA Dialect (tosa.py)

• Implemented 171 new operator lowering functions.
• Math ops: abs, log, ceil, floor, log10, log2, log1p, exp2, expm1, sqrt, rsqrt, sign, digamma, lgamma, i0, etc.
• Comparison ops: eq, ne, gt, ge, lt, le (both Tensor and Scalar variants).
• Logical ops: logical_and, logical_or, logical_xor, logical_not.
• Bitwise ops: bitwise_not, bitwise_and, bitwise_or, bitwise_xor.
• Reduction ops: sum, prod, mean, std, var, amin, amax, argmax, argmin, cumsum, cumprod, all, any, norm.
• Pooling ops: avg_pool1d/2d/3d, max_pool1d/3d, adaptive_avg_pool1d/2d/3d, adaptive_max_pool1d/2d.
• Activation ops: gelu, elu, selu, celu, leaky_relu, hardsigmoid, hardswish, hardtanh, hardshrink, softshrink, softplus, mish, prelu, rrelu, threshold.
• Tensor manipulation: reshape, squeeze, unsqueeze, permute, transpose, expand, slice, select, gather, scatter, flip, roll, tile, unfold, narrow, split_with_sizes.
• Normalization: native_layer_norm, native_group_norm, and native_batch_norm variants.
• Other: upsample, pad, clamp, where, masked_fill, sort, topk, embedding, diagonal, etc.

Linalg Dialect (linalg.py)

• Added 25 new operator implementations.
• Restored special broadcast pattern handling in index_op for cases like idx0: (1, 1) + idx1: (N,).
• Fixed ops_registry mappings for LtTensorOp and SqueezeDimOp.

Math Dialect (math.py)

• Added implementations for sinh, cosh, tan, erf, exp2, and other math operations.

---

Test Coverage

• Added 198 new test files in tests/Python/AtenOps.
• All tests run with DynamoCompiler and use FileCheck for IR validation.
• The new tests provide comprehensive coverage for all newly implemented operators.

---

Known Issues

In PyTorch 2.8.0, the interaction between AOTAutograd and Aten IR still has some limitations. Under the default inductor_decomp decomposition rules, certain operators cannot be lowered correctly and require custom decomposition settings:

• tests/Python/AtenOps/test_reflection_pad1d.py must set aot_autograd_decomposition=None.
  Otherwise, the Inductor decompositions produce a complex sequence of
  tosa.const + tosa.abs + tosa.sub + linalg.generic + tensor.extract,
  even though aten.reflection_pad1d is marked as a Core Aten IR operator in the documentation.

• In tests/Python/AtenOps/test_max_pool3d_linalg.py, max_pool3d must use core_aten_decompositions() instead of the default inductor_decomp rules.
  The Inductor decomposition for max_pool3d triggers a tracing error, while core_aten_decompositions() keeps the operator intact so that buddy-mlir can directly use its own implementation.

These issues are limitations of the current PyTorch 2.8.0 decomposition behavior rather than the backend changes in this PR.