[Feature Request] Dataset cache for COCO annotations that don't fit in memory

## 🚀 Feature
I'd like to request adding [D2Go's cache feature](https://github.com/facebookresearch/d2go/blob/87374efb134e539090e0b5c476809dc35bf6aedb/d2go/data/config.py#L39) to the main Detecton2 repo.

## Motivation & Examples

This feature would make it easier to train on very large datasets.

Currently it's extremely cumbersome to train and perform coco evaluation on a custom format other than coco (for example a TorchData iterable), though coco format annotations may not fit into memory for very large datasets.

Start by reading the linked D2Go cache implementation in d2go/data/config.py, then trace Detectron2's dataset loading and COCO evaluation paths for where annotations are held in memory. Done means large COCO-annotation datasets can be cached and used for training and evaluation without requiring all annotations to fit in memory.

Written by the indexing model from the issue text.