Support Float16

[mtbarta]

Since late interaction is already so expensive storage-wise, we can support float16 as a quick win to reduce storage costs.

This would involve adding support for a new datatype. Alternatively, we could consider a quantizer that converts float32 to float16. I'm preferential to a new data type.

[mtbarta]

Let's break this idea up a bit.

We should support float16 for individual FieldValues. For Tensors, quantization should be preferred in reducing storage size.

Also note, we should support float16 computation as well. We can break that out into a separate ticket.