--precision 16 crashes on Segtrain with CPU

[PonteIneptique]

FYI, I'd probably add a sanity check here:

RuntimeError: expected scalar type BFloat16 but found Float

(Note that I just forgot to add --device cuda and that's how I found out)

[colibrisson]

Maybe we can add a line in the training advice to encourage people to use mixed precision.

[PonteIneptique]

If device CUDA -> "You can optimize training by using precision 16" ?

[colibrisson]

Do you mean to print a warning when training on GPU using 32-bit precision? I was just thinking about adding a line to the documentation.