half and mixed precision inference

[ArneBinder]

This PR implements half and mixed precision inference by adding the following boolean parameters to the Pipeline:

• half_precision_model: Whether or not to use half precision model. If set to True, the model will be cast to half precision on supported devices (torch.float16 on cuda and torch.bfloat16 on cpu for now, following torch.get_autocast_dtype()). This can reduce the memory consumption and improve the inference speed, but may lead to numerical instability.
• half_precision_ops: Whether or not to use half precision operations. If set to True, the model will be run with half precision operations via torch.autocast.

Since torch.float16 can not be natively converted to numpy, this PR adds float() casting to several model outputs (probabilities) before calling .numpy(). in unbatch_output() of several taskmodules. Also, this fixes some spelling mistakes in the Pipeline documentation.

TODO:

• test in downstream setup, e.g. inference on drugprot.
  • Outcome: effective for larger batch sizes (>1)

batch_size	half_precision_model	half_precision_ops	inference_time	speed_up
32	True	True	32.8	2.4
32	True	False	26.4	3.0
32	False	True	28.8	2.7
32	False	False	79.0	1.0
256	True	True	23.6	3.5
256	True	False	21.2	3.9
256	False	True	23.7	3.5
256	False	False	83.6	1.0
512	True	True	24.4	3.7
512	True	False	22.7	4.0
512	False	True	25.0	3.6
512	False	False	90.7	1.0
1024	True	True	26.9	3.7
1024	True	False	23.7	4.1
1024	False	True	26.7	3.7
1024	False	False	98.3	1.0