feat: process images

[qued]

Adds ability to process images as single page documents. Uses mostly the same path as processing a pdf, but uses a null layout (since an image won't provide a pdf layout) and does not filter layout text blocks using a bounding box to discover text (because there are no layout text blocks).

Testing:

Run this code in the unstructured-inference environment:

from unstructured_inference.inference.layout import DocumentLayout

doc = DocumentLayout.from_image_file('sample-docs/loremipsum.png')
print(doc.pages[0].elements)

[MthwRobinson]

LGTM! Longer term, curious if there are cases where it may make sense to go straight to PaddleOCR instead to save on compute time. This looks great for the detectron2 case.