remove cast to float32 and float16 when using CPU for pushing to HF

[pascal-pfeiffer]

I think recent changes in 3rd party dependencies make this cast to float32 and float16 obsolete and CPU seems to be able to handle it.

closes #702

[pascal-pfeiffer]

I found this as an explanation why the cast to float32/16 was added.
Wdyt, shall we just remove and see when it breaks or wrap in a try/except block? Ideally, we will add a selection for target precision (or quantization) anyway.

[psinger]

I found this as an explanation why the cast to float32/16 was added. Wdyt, shall we just remove and see when it breaks or wrap in a try/except block? Ideally, we will add a selection for target precision (or quantization) anyway.

yes I think we can see when it breaks