Changes to parsing behavior, documentation and tests

[dtronmans]

Purpose

• Allow keypoint annotations parsing for COCO using the use_keypoint_ann when test annotation json is
• Revert previous change: previously I allowed all images to contain hard negatives (no ground truth labels) for dense estimation tasks (instance segmentation, detection and keypoints). Now just for coco-2017 keypoints we don't parse these images with no annotations, because over 50% of the images in the dataset don't have keypoints
• Documentation regarding parsing coco-2017 and imagenet-sample to make everything clearer
• Test structure and first few tests for parsing using CLI, we can add more over time depending on edge cases that appear or behavior we want to enforce.
• Added coco keypoints to test_parse_export_equivalence and added equivalency testing for semantic segmentation masks (realized I had it for instance segmentation but not semantic segmentation)

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation

[klemen1999]

One small nitpick but overall LGTM

[ptoupas]

Left a couple of minor comments, but apart from that it LGTM.