Feature description

Add support for f16/bf16 for the ndarray backend, mostly through the half::f16/bf16 types.

Feature motivation

• Some other backends support it, such as Cuda. This could also be interesting for wasm targets, given that they can be edge devices.
• For debugging purposes, in case an architecture has stability issues in f16, it's desirable to able to debug it in ndarray than in an gpu-backed device.

Suggest a Solution

The natural way forward is for ndarray to support f16 types. I've opened an issue for it rust-ndarray/ndarray#1550, and I plan on having a draft/concept PR for this as a upstart (for now, relying on patches of ndarray). I believe this requires minor changes to both ndarray and burn.