`Tensor.grad_check` does not evaluate many supported operations

[github-actions]

File: leanpass/tensor.py

The helper _eval_forward used by grad_check only implements a subset of ops (+, -, *, /, **, @, relu, exp). Operations such as log, sigmoid, softmax, tanh, gelu, clip, etc., are missing, causing grad_check to raise NotImplementedError or produce incorrect numeric gradients for models that use these functions. Extend _eval_forward to handle all ops defined in Tensor or document that grad_check is limited to the supported subset.

Label: enhancement

Filed automatically by ai-issue-scan.