Inference improvements

[mortonjt]

On inference time, we are currently not batching. So there is room for improvement for speeding up the featurize_protrans method. I anticipate that we can get a 10-100x improvement in speed. Batching is already being done on training these models, so it is a matter of porting over some of the existing code. See below for sorting the priority of tasks that can boost performance (in terms of low hanging fruit)

• Batching on protrans models (>10x)
• Batching on tm-vec embeddings
• xformers (>4x)
• torch.compile (>2x)
• Precision (16 bit precision, 8 bit precision) (>2x)
• Multiple GPUs (8x)
• Model distillation of Protrans

Relevant issue : flatironinstitute/deepblast#130