Rescale TGLFNN predictive variances correctly

[sylvesterkaczmarek]

The TGLFNN output's mean and variance currently share the same affine unnormalization. Adding the output mean to the variance can produce negative uncertainties; scaling by the standard deviation also gives the wrong units.

Keep mean unnormalization unchanged, but give the variance a zero offset and squared standard-deviation scale. Preserve the existing zero-standard-deviation fallback and the output dtype.

Validation

Full CPU suite: 107 tests passed. All 12 new regression cases fail on unpatched main. The tests use a small real GaussianMLP ensemble with known mean and variance, covering scalar/batched inputs, float32/float64 outputs, JIT, and zero-standard-deviation handling. Pyink, isort, and git diff --check pass. No GPU/TPU execution was performed.

[sylvesterkaczmarek]

Could a maintainer approve the workflow and review the TGLFNN predictive-variance rescaling fix? The change corrects the variance transformation when converting predictions back to the output scale.