[Fix][Relax][Torch] Respect zero-dimensional tensor dtype promotion

[jack-k737]

Problem

PyTorch dtype promotion depends on tensor rank. A zero-dimensional float64 tensor added to a dimensioned float32 tensor produces float32, but the Torch importer used torch.promote_types and incorrectly promoted the result to float64.

Changes

• Use torch.result_type on meta tensors when both operand ranks are known.
• Preserve the existing torch.promote_types behavior for unknown ranks.
• Add regression coverage for both operand orders.

Testing

• python -m pytest tests/python/relax/test_frontend_from_exported_program.py -k 'binary_dtype_promotion' -q (8 passed)
• pre-commit run --files python/tvm/relax/frontend/torch/base_fx_graph_translator.py tests/python/relax/test_frontend_from_exported_program.py

[jack-k737]

One-line summary

The legacy TVM export frontend derives an incorrect dtype for binary operations involving zero-dimensional (0-d) tensors: torch.Tensor(1.0, dtype=torch.float64) + torch.Tensor([1.0], dtype=torch.float32) actually returns float32, whereas the export frontend infers float64 based on the legacy promotion rule.

Detailed description

In legacy TVM releases, the export frontend exhibits a type-promotion discrepancy for binary operations involving 0-d tensors (e.g., torch.Tensor(1.0)). At runtime, PyTorch treats 0-d tensors as scalars (wrapped numbers) under weak type-promotion semantics: in the expression above, the float64 0-d operand defers to the float32 1-d operand, so the result is float32. The export frontend, however, does not account for the scalar semantics of 0-d tensors and falls back to the legacy strong tensor promotion rule (float64 takes precedence), inferring the output dtype as float64. This mismatch between compile-time type inference and runtime semantics leads to dtype inconsistencies during golden/accuracy validation of the exported graph or in downstream compilation.

Root cause: The export frontend's type inference does not apply PyTorch's scalar semantics to 0-d operands and still relies on the legacy promotion rule.

Suggested fix: Infer the types of 0-d tensor operands using PyTorch's weak promotion rule so that frontend inference aligns with runtime behavior.