Missing BFloat16 support #2573
Description
We should have also bf16 besides f16, f32, and f64.
The package BFloat16s could be a dependency or support could be provided by an extension.
Reactant seems to support BFloat16 (although we don't support reactant yet, but things may change soon FluxML/Fluxperimental.jl#28)
Support in CUDA.jl is incomplete, see JuliaGPU/CUDA.jl#2441