[LauraGPT]

Summary

Integrates #3853 while preserving the original contributor commit and authorship. The appended signed commit synchronizes retained main content to the exact tested integration tree without rewriting history. UtteranceMVN's both-enabled branch excludes centered padding from variance and divides by standard deviation once. Other normalization branches retain their existing behavior.

Adds the submitted regression file to NumPy CI. The audio integration test now explicitly uses IEEE FP32 rather than environment-dependent TF32, restoring both backend settings afterward. AST comparison verifies every original assertion, tolerance, reference and test body is unchanged.

Maintainer validation

• On unchanged main: 21 of 55 submitted CPU/CUDA cases fail, including a preexisting TF32-sensitive mean-only encoder comparison. Candidate: 104 pass and 2 CUDA audio comparisons fail out of 106 selected cases; that failure is retained.
• Falsifiable precision check: with TF32 disabled only in the test process, unchanged main still fails the variance-normalized audio case (max valid-feature difference 2.441162); the candidate passes both CUDA cases at unchanged tolerances.
• The added fixture passes all 4 CPU/CUDA audio cases, 12 real inference calls, 629 CUDA kernels. The 102 unaffected selected cases are retained against unchanged source, giving 106 unique accepted selected cases. No full-suite claim.
• Real L20Z checks cover FP64/FP32/FP16/BF16 normalization, padding/order/batch invariance, epsilon, gradients, unchanged switches, local generated checkpoint round-trip, and repository audio through a small randomly initialized Transformer. These are not pretrained recognition-quality results.
• Private HOME/cache and kernel network isolation; no new pretrained weights or shared/production environment changes. Exact new-head hosted CI is required before normal merge; hosted CPU CUDA skips do not replace L20Z evidence.

Compatibility

Both norm_means=True and norm_vars=True now produce corrected feature values. Existing checkpoints trained with both enabled require recognition-quality reevaluation. norm_vars defaults to False; this PR does not claim pretrained quality is unchanged. No release, Nano work or production modification is included.