[Relax][FRONTEND][ONNX] Support Softmax, LogSoftmax and Hardmax when opset version ≤12

[cchung100m]

Hi Commiters,

This PR is going to fix Softmax-family legacy semantics for opset<=12 and harden Hardmax fallback.

Summary:

• Implement legacy ONNX semantics for Softmax / LogSoftmax / Hardmax in opset <= 12, including flatten-to-2D + reshape-back behavior.
• Keep opset >= 13 behavior unchanged (_impl_v13 axis-based path).
• Add compatibility-first fallback warnings for unknown rank/shape paths.
• Harden Hardmax internals: normalize input before struct_info access and add backward-compatible helper signature handling.
• Extend regression coverage for softmax-family across opset 1/11/13 and key axis scenarios.

[gemini-code-assist]

Code Review

This pull request implements legacy semantics for ONNX Softmax, LogSoftmax, and Hardmax operators (opset <= 12), which require flattening the input tensor into a 2D matrix based on the axis attribute. The review feedback correctly identifies that the current fallback logic for tensors with unknown static rank incorrectly defaults to modern (opset >= 13) behavior. It is recommended to raise a ValueError in these cases to prevent silent calculation errors for dynamic-rank tensors.

[cchung100m]

Hi @tlopex @mshr-h

This PR is trying to fix issues #19420. Any suggestions would be appreciated if you are available.

[tlopex]

LGTM