Feat/chipper gpu float16

[ajjimeno]

We added the automatic mixed precision for CPU, but this is needed as well for GPU running. This PR makes this possible for GPU.

For testing, probably before and after the change on any long page.

from unstructured_inference.inference.layout import DocumentLayout
from unstructured_inference.models.base import get_model

model = get_model("chipper")
doc = DocumentLayout.from_file("sample-docs/layout-parser-paper.pdf",
                               detection_model=model, pdf_image_dpi=300)

[LaverdeS]

LGTM. Tested on colab (login in to HF to use chipperv2) with a T4 using the snippet:

%%time
doc = %memit -i3 DocumentLayout.from_file("donut_rl_merged.pdf", detection_model="chipper", pdf_image_dpi=300)

Here are the results for unstructured==0.7.10:

peak memory: 5038.44 MiB, increment: 53.56 MiB
CPU times: user 16min 18s, sys: 34.1 s, total: 16min 52s
Wall time: 16min 36s

And the results for unstructured==0.7.10-dev0 (with the changes in this PR):

peak memory: 4999.92 MiB, increment: 940.64 MiB
CPU times: user 17min 19s, sys: 36.8 s, total: 17min 56s
Wall time: 17min 37s

Both version executed a parsed a pdf document with 66 pages (academic paper) successfully.

There is a big change in the memory increment value though, from 53.56 MiB to 940.64 MiB, but the GPU RAM looks better. In the image, the first part of the GPU RAM region until the red top ends, corresponds to unstructured==0.7.10, then there is a runtime restart, and the rest in orange is with unstructured==0.7.10-dev0.