mnist demo improvements

[ajsenthi]

MNIST Demo Improvements — accuracy & inference improvements

Fixes several bugs degrading training accuracy and canvas inference quality in the MNIST federated learning demo.

store_data.py

Stratified data loading
Replaced sequential loading with per-class stratified sampling (both train and test sets) to eliminate label bias from MNIST's sorted order.

custom_data_handler.py

Training pipeline
Normalization: moved /255.0 into load_dataset(); preprocess() now only casts dtype.
Query: removed WHERE round_number filter — each round now trains on all available data (LIMIT 1000 train / LIMIT 200 test).
Regularization: added Dropout(0.25) after each conv block and Dropout(0.5) before output.
Class weights: compute_class_weight('balanced') applied per round to handle class imbalance.
Hyperparameters: batch size 128→32, epochs 1→5, EarlyStopping patience set to 5.
Inference endpoint: direct_inference() now normalizes raw 0–255 input automatically.
Evaluation: replaced val_accuracy with run_inference() against TEST_TABLE after each round.

InputDataSelector.js

Canvas brush
Changed from 1×1 pixel brush to a 5-cell cross pattern, better matching MNIST stroke width.

InferPage.js

Canvas preprocessing
Added centerAndScale(): fits drawn digit into 20×20 region, centered on 28×28 grid.
Added gaussianBlur(): 3-pass blur + max-normalization to smooth binary canvas output.

Test results

JSON (normalized 0–255), 30-sample set: ~92%, originally was ~68%, about 24% improvement.
Canvas — all digits 0–9 ✓, testing done by hand.