Fix ContextBuilder checkpoint loading for non-default architectures

[harens]

This PR updates ContextBuilder persistence so checkpoints store the architecture metadata needed to reconstruct the model reliably.

Previously, save() wrote only the raw state_dict, and load() inferred constructor settings from tensor shapes. That worked for some default models, but was unreliable for non-default configurations such as custom num_layers, bidirectional encoders, or LSTM-based models.

Changes:

• Store ContextBuilder constructor metadata alongside the state_dict.
• Restore saved input_size, output_size, hidden_size, num_layers, max_length, bidirectional, and LSTM values on load.
• Preserve backwards compatibility with older raw state_dict checkpoints.
• Infer num_layers, bidirectional, and LSTM from recurrent tensor keys/shapes where older checkpoints do not include metadata.

This should make checkpoint round-tripping reliable for non-default architectures while keeping existing saved models loadable.

For legacy checkpoints, architecture parameters are inferred from PyTorch recurrent weight naming conventions (e.g. weight_ih_l{k},_reverse suffix, and gate dimensionality). This is a best-effort heuristic.