[faildeny]

Fix NN uncertainty evaluation and UncertaintyWeighted aggregation

Changes

• Binary MC-dropout metrics fixed. predict_proba_mc applied softmax to a single logit, so binary models always reported entropy 0 and an accuracy equal to the share of class 0. It now uses sigmoid. evaluate() always reports the standard metrics and adds the MC-dropout ones (entropy, val_accuracy) only for classification.
• Saved NN models load in inference.py. The model metadata now includes dropout_p and mc_samples (= T). Without them load_model refused to rebuild the model.
• n_out is consistent across sites. load_dt4h no longer replaces the metadata class count with the number of classes present locally. A site missing a class used to get a different output layer from the other sites.
• UncertaintyWeighted aggregates evaluation metrics. The strategy now receives evaluate_metrics_aggregation_fn, so metrics_distributed is populated.
• Regression falls back to FedAvg. Clients only report entropy for classification, so UncertaintyWeighted with a regression task now switches to FedAvg at config time.
• accept_failures is respected in UncertaintyWeightedFedAvg.aggregate_fit.
• Tests: a new test_nn_uncertainty_weighted covers binary, multiclass and regression with MC dropout on. fl_runner gains server_extra_args. CI also runs on PRs to tests_and_CI.

New server arguments (server_cmd.py)

Argument	Default	Effect
--uw_epsilon	0.1	Smoothing term in the client weight n / (ε + entropy)^γ. Limits how much a client with near-zero entropy can dominate (max ratio ≈ (1+ε)/ε).
--uw_gamma	1.0	Strength of the uncertainty weighting. 0 is plain FedAvg; larger values give more weight to low-entropy clients.

Both only apply with --strategy UncertaintyWeighted.