Fix efficientnet-lite0 fine-tuning regression from the timm backbone swap

[gomezzz]

Summary

Switching the efficientnet-lite0 backbone to timm (tf_efficientnet_lite0) silently dropped three model-level defaults that the previous efficientnet_lite_pytorch package applied. Together they regressed FixMatch fine-tuning quality substantially. This PR restores all three:

• Classifier head over-scaling (main cause). timm initialises the head with TF-EfficientNet's 1/sqrt(fan_in+fan_out) scale, tuned for the 1000-class ImageNet head — roughly num_classes too large for AnomalyMatch's 2-class head (weight std ~0.4 vs PyTorch's ~0.016). The overconfident fresh head makes ~100% of unlabeled images clear the FixMatch p_cutoff (0.95) from the very first step with essentially random pseudo-labels, poisoning the backbone. Fix: reset the classifier to PyTorch's default init after timm.create_model(...).
• cfg.bn_momentum never applied. It is computed (= 1 - ema_m, ~0.01) and validated but never set on the model, so BatchNorm ran at timm's 0.1 default (~10x too fast for our small batch size). Fix: apply it to both train_model and eval_model BatchNorm layers.
• cfg.seed never applied. The seed was defined and validated but set_seeds was never called, so training was not reproducible. Fix: seed all RNGs after config validation, before datasets and model are built.

A regression test asserts the classifier head weight std stays near PyTorch's Linear bound.

Validation (internal benchmark, AUROC)

Class	Before (timm regression)	After fix
hourglass	0.81	0.94 – 0.96
piano	0.83	0.89 – 0.92

Test plan

• ruff check + ruff format clean on changed files
• New tests/unit/test_net_builder.py::test_efficientnet_classifier_head_init_is_pytorch_scale passes
• tests/unit/test_fixmatch.py and tests/unit/test_import.py still pass
• anomaly_match.pipeline.session imports cleanly

[gomezzz]

Superseded by #16, which uses a direct branch on this repo (no fork) — this clears the fork-only coverage-comment CI restriction. Closing in favour of #16.