Fused operations

[bclarkson-code]

A lot of time is spent in MLP blocks and attention blocks. The operations in these blocks can be fused to reduce both memory usage and latency