Fix degenerate quantile normalization

[milkieTigera]

Summary

• Fall back to z-score normalization per dimension when q99 - q01 <= 1e-6.
• Apply the same fallback during unnormalization so the transform remains symmetric.
• Preserve action dimensions that do not have normalization statistics.

Motivation

Constant or nearly constant dimensions can have a degenerate quantile range. Applying quantile scaling to those dimensions amplifies tiny numerical differences into very large normalized values. The existing mean and standard deviation statistics already provide a stable fallback for those dimensions.

Compatibility

The change is limited to degenerate dimensions when quantile normalization is enabled. Valid quantile ranges continue to use the existing formula, and the default z-score path (use_quantiles=False) is unchanged. No model, data loader, or policy configuration is modified.

Validation

• Targeted local regression tests: 5 passed.
• Red/green check: the original implementation failed the two fallback-specific cases; the patched implementation passes all five cases.
• Python compilation check passed for the modified module.
• git diff --check passed.