Revert sklearn to 1.5.1, fix multi_class at load time

LSTM: fallback to rebuild model arch and load weights

LSTM: try keras.saving.load_model first

LSTM: use native keras (Keras 3) for loading, remove tf-keras

LSTM: use keras3 directly (not tf-keras) for loading Keras 3 model

LSTM: detect file format header, better diagnostics

LSTM: detailed loading diagnostics

Add load_errors to health endpoint for debugging

Fix LSTM: better loader fallbacks, verify tf-keras

Fix LSTM loading: try compile=False and multiple backends

Fix: resilient predictions, sklearn 1.3.2, LSTM compat, CPU torch

Pin sklearn to 1.3.2 for model compatibility

Fix sklearn compatibility for LR model (multi_class attribute)

Initial commit: Sentiment Analysis HF Space