Fix aten::mean dtype handling

[Promise Emmanuel Oluwadare (promiseeuler)]

Summary

While testing ONNXScript's TorchLib lowering for torch.mean(..., dtype=...), I found that the aten::mean overload ignored its schema's dtype argument and silently exported the wrong output type.

This change:

• accepts the optional dtype argument
• casts the input before ReduceMean, preserving the requested accumulation precision
• adds a precision-sensitive end-to-end regression

Fixes #3008.

Validation

• Exported torch.mean([1e8, 1, -1e8], dtype=torch.float64)
• Verified graph order: Cast -> ReduceMean -> Squeeze
• Verified ONNX reference output matches PyTorch in value and float64 dtype
• ruff check passed
• ruff format --check passed

The focused e2e test file cannot collect on this local Python 3.14 build because torchvision imports the unavailable stdlib _lzma extension; the same test logic was executed directly and passed.

[Promise Emmanuel Oluwadare (promiseeuler)]

Thanks om singhal (@Om-singhaI) — I found #3009 after opening this PR. I compared the two changes, and #3009 covers the same cast-before-reduction fix with broader OpInfo coverage and consistent handling of the complex overload. To keep maintainer review focused and avoid duplicate work, I’m closing this one in favor of #3009. The precision-sensitive reproduction from #3008 agrees with its approach.