Fix for type mismatch bug and numerical stability improvements

[shantanusharma]

This PR addresses a type mismatch bug in eigenvalue merging and improves numerical stability in Lennard-Jones force calculations.

Key changes:
- In `torch_sim/math.py`: Fixed a `TypeError` in `_determine_eigenvalue_case` where boolean masks were incorrectly combined with integer tensors. Added proper device handling for `torch.arange`.
- In `tests/test_math.py`: Added comprehensive tests for `_determine_eigenvalue_case` and `_matrix_log_33` with numerically close eigenvalues to prevent regressions.
- In `torch_sim/models/lennard_jones.py`: Implemented safe division in `unbatched_forward` to avoid division by zero when particles overlap (distances = 0).
- In `torch_sim/transforms.py`: Added a check to `inverse_box` to raise a `ValueError` if the box contains zero values, preventing division by zero.

Checklist:
- [x] Doc strings have been added in the Google docstring format.
- [x] Run ruff on your code.
- [x] Tests have been added for any new functionality or bug fixes.