[han-xudong]

Summary

• add variadic Max and Min with multidirectional broadcasting for rank-0, rank-1, and rank-2 policy tensors producing rank-2 outputs
• preserve gradient propagation and allocation-free CUDA graph replay through preallocated intermediate buffers
• document the newly supported operators

The branch is rebased onto the develop version that includes #8, and the duplicate Tanh implementation has been removed.

Testing

• .venv/bin/python -m pytest tests/runtime/test_onnx_runtime.py -q (14 passed)
• uvx pre-commit run -a

Closes #6

[Toni-SM]

Hi @han-xudong

Could you please rebase and solve merge conflicts?
Also, please note that Than operator was added in #8.

[han-xudong]

@Toni-SM Thank you for the guidance. I rebased the branch onto the latest develop and resolved the conflicts with #8. I removed the duplicate Tanh implementation and narrowed both #6 and this PR to the remaining variadic Max and Min support. The complete ONNX runtime test file passes (14 tests), and all pre-commit checks pass.