Support pre-packed weights

[robertknight]

This reduces inference time for matmul operations at the cost of higher memory usage. Prepacking is initially disabled by default as there are common situations (single inference run, large transformer decoder models) where it isn't worthwhile. It is enabled via ModelOptions::prepack_weights:

let mut model_options = ModelOptions::with_all_ops();
model_options.prepack_weights(true);
let model = model_options.load_file(path)

To facilitate a quick comparison of the performance impact of prepacking, the rten CLI now has a --prepacked option to enable prepacking.

There are some caveats:

• Non-MatMul operations which use matmuls internally (Conv, ConvTranspose, LSTM, GRU etc.) currently don't prepack their weights.
• MatMul operations which turn out to be matrix-vector (gemv) products don't use the prepacked weights. This affects transformer decoders doing non-batched generation after the initial prompt encoding step.

Fixes #214.

TODO:

• Tests
• Make it configurable whether prepacking is used
• Support subgraphs