Enhance AssignResult and SamplingResult

[Erotemic]

Added random classmethods to SamplingResult and AssignResult.

These allow for the simple creation of "demo data" that can be use for testing, debugging, and potentially more (although I don't think these are as useful as random boxes might be).

Fix issue in SamplingResult.__init__

I've been hitting issues where sometimes gt_bboxes looks like its had its dimensions squashed. This causes an error when doing the index selection. Adding two checks fixes it.

Add rng as attribute of RandomSampler

To ensure that the new random methods could be seeded to consistently reproduce an issue if needed, I had to ensure that the RandomSampler class could be seeded. I used ensure_rng, which I added in a previous PR to accomplish this simplify.

Misc

I also added a to function to SamplingResult, where it moves the data onto a different device. I wrote this to test something that turned out not to be an issue, but I figure its a nice API to have so I left it in.

I also added info properties to SamplingResult and RandomSampler as well as implementing the __nice__ debugging string for SamplingResult.

I did end up using my ubelt library to accomplish some of these tasks. The core fix and the random clasmethods don't rely on ubelt, so if there is any resistance to adding a dependency I can remove the features that depend on ubelt and still have a functionally useful PR. However, I've been maintaining and keeping ubelt stable for ~2 years, it has 100% test coverage, it itself is fairly small, and its own dependencies are minimal.

[hellock]

Thanks for the PR as well as previous ones! ubelt is indeed an excellent util library, but it may be too early to introduce it since we only need two methods/classes of it. We can redefine __repr__ and __str__ to achieve the same goal of NiceRepr.

[Erotemic]

Sounds good. I removed ubelt as a dependency.

Instead of redefining __repr__ and __str__ (which imo would unnecessarily bloat the code) I simply copied the definition of NiceRepr into mmdet.utils. This removes the dependency on ubelt, but keeps the succinct and easy-to-modify __nice__ definition. It also prevents code redundancy in both SamplingResult and AssignResult.

I was also using ubelt.repr2 in this PR, which unfortunately is a bit too complex to simply copy-paste into a code base, so I simply wrote out some logic that produces a similar effect. Its less pretty than having a simple function that nicely handles nested representations, but it does prevent the extra dependency.