fix: prevent torch checkpoint offset overflow

[leejet]

Summary

• Reject negative or overflowing PyTorch checkpoint storage sizes, tensor offsets, and tensor layouts.
• Replace wrap-prone storage bounds checks with overflow-safe subtraction checks for zip and legacy checkpoints.
• Revalidate zip entry bounds immediately before copying tensor data and propagate entry read failures.

Related Issue / Discussion

N/A

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.