Feature subsample

[clegaard]

Summary:

1. Added subsampling transform which allows the user to "divide" each sample into one or more smaller sub-samples.

The current implementation requires the user to specify the sub-samples produced by each sample.
Which also limits it to a subsampling process that always produce a fixed number of subsamples.

2. Updated documentation on this.

[LukasHedegaard]

@clegaard , now that the caching branch has been merged into develop, there are some merge conflicts that need fixing before the merge. Also, more severe linting checks were added, so the branch needs to be checked for these.

[LukasHedegaard]

Hey, a couple of things:

1. Cool that you remembered to include a test for caching! 👍
2. Your commit naming doesn't follow our standard (https://chris.beams.io/posts/git-commit/) - please use it in the future. And what kind of commit message is more push? 😂
3. You have snuck in the csv loader, I see - how about canceling the other PR and just include it here?
4. You're returning a named tuple "Row" from the csv loader - this doesn't comply with our internal structure and should be changed to a regular tuple. If you believe you can infer names, how about setting the dataset names instead?
5. The naming of internal variables should start with '_' if they are not supposed to be directly used by the end-user. Most of the internal variables, you introduce don't follow this convention. Moreover, you've made some of them a tad brief: n_ss for instance should be written out to make it easier to understand.
6. It seems there is a bit of confusion in the class implementation. For instance, SubsampleDataset takes a dataset parameter and saves it a self.dataset. This is exactly the same as _downstream_getter (now _parent) used in the regular Dataset, so why save two versions?
7. I'm not that big a fan of overloading the dataset. You get a lot of extra state keeping, which you also expose to the user here. I think we should at least wrap the SubsampleDataset returned from subsample in another Dataset to limit the exposure of our guts.

Aside from this, I noticed and fixed some typing problems and snuck in a few issue fixes:
#41
#31

Also, I've renamed cachable (which was misspelled) to _cacheable (assuming we want it to be private)

[LukasHedegaard]

Hey, slightly better commit names. However, you're still not following the standard (did you actually read the document?). The names should be:
"Change attribute names and remove dataset reference"
"Change from_csv to return a plain tuple"

[LukasHedegaard]

I can see, you have added slicing. While it seems to work for Dataset, it doesn't for all the child classes that overload __getitem__ (StreamDataset, SubsampleDataset, SupersampleDataset). Moreover, the ItemGetter should also be updated to take slices if you want to be able to generically use self._parent[1:2] within the dataset class