Relax Test Tolerances for TE executor tests

[riccardofelluga]

Summary

This PR introduces dtype-specific numerical tolerances for Transformer Engine (TE) tests to improve test stability and reduce false positives due to numerical precision differences across different floating-point data types.

Key Changes

• Added te_assert_close function: A new assertion wrapper that automatically applies appropriate tolerances based on tensor dtype and scalar inputs
• Dtype-specific tolerances: Implements tolerances aligned with Transformer Engine's numerical testing specification:
  • float32: rtol=1.3e-6, atol=1e-5
  • float16: rtol=1e-3, atol=1e-5
  • bfloat16: rtol=1.6e-2, atol=1e-5
  • float scalars: rtol=1.3e-6, atol=1e-5

[mattteochen]

lgtm, thank you!

[t-vi]

Thank you @riccardofelluga @mattteochen