Fix/reduce memory usage

[badGarnet]

This PR fixes the high memory usage when computing intersection areas.

• it now converts the coordinates into half precision floating point numbers instead of double
• removes some intermediate variables to free up memory usage

test

Using a memory profiler like memory_profiler in ipython:

## cell 1
from unstructured.partition.pdf_image.pdfminer_processing import areas_of_boxes_and_intersection_area
import numpy as np
%load_ext memory_profiler

## cell 2
%%memit
coords = np.random.rand(40000).reshape((10000,4)).astype(np.float16)

## cell 3
%%memit
inter_area, boxa_area, boxb_area = areas_of_boxes_and_intersection_area(coords, coords)

The peak memory and incremental memory from cell 3 should be close to

peak memory: 730.55 MiB, increment: 573.22 MiB

On main branch the coords is double precision and running the same code with

coords = np.random.rand(40000).reshape((10000,4)).astype(np.float64)

would result in peak memory usage more than 4GiB

[christinestraub]

There are some unusual changes in the ingest test fixtures. I'm investigating this issue.

[christinestraub]

LGTM!