support multiple sampling methods, instead of random sampling only

Nobody has claimed this yet.

## 🚀 Feature
Many new object detection algorithms proposed some novel sampling methods. I hope detection2 could provide a BaseSampler class which could be easily extended into other sampling methods.

## Motivation & Examples
For instance, LibraRCNN proposed an IOU-Balanced sampling method and [PISA](https://arxiv.org/pdf/1904.04821.pdf) also focused on the sampling approach. Experiments prove that using these new methods lead to an obvious improvement. A set of sampling methods have been integrated into [mmdetection](https://github.com/open-mmlab/mmdetection/tree/master/mmdet/core/bbox/samplers)

## References
[Prime Sample Attention in Object Detection](https://arxiv.org/pdf/1904.04821.pdf)
[Libra R-CNN: Towards Balanced Learning for Object Detection](https://arxiv.org/pdf/1904.02701.pdf)

Start by reviewing the sampling implementations in mmdetection's mmdet/core/bbox/samplers directory and the linked PISA and Libra R-CNN references. Determine how detectron2 could expose an extensible BaseSampler for alternatives to random sampling; done means multiple sampling methods can be supported through that extension point.

Written by the indexing model from the issue text.