Ensures source and target IDs are paired up in SGD

[weefuzzy]

fixes #292

Ok, ok, I know this looks like a massive PR, but it's not really 😄

The problem

SGD::train assumes that its source and target data are aligned, i.e. that each index represents the same ID. This not necessarily the case because we're using unordered containers by design. You can make training silently fail to do anything sensible by adding points to source and target in a different order, for instance. This is baaaaad.

But why is this diff so large, you 'orrible man?

The core of this PR is decoupling the batching and sampling of data from SGD, a bit like pytorch. I like that SGD doesn't want to know about DataSet, it just needed to go further.

So we have

• some new classes that do data sampling: a base class, an implementation that properly matches source and target IDs, and a simple implementation for when data really are aligned (this opens up some cool future extensions*).
• changes to SGD to remove its batching code and use this instead
• supporting changes to FluidDataSet, FluidTensorView
• some tests
• updates to MLP clients to actually check that IDs match and use appropriate sampler

User-side, nothing much should change besides a couple of error messages: so, reviewers: please test this proposition by training against your own stuff and making sure it still works as it did.

* Like, it would now be pretty easy to enable quicker n' dirtier MLP training (resp. the rest of the data objects) with CCE buffers alone for those cases where people don't want / care about IDs. I think this would have significant workflow and teaching upsides...

[tremblap]

Thanks for this mammoth work! Maybe @lewardo wants to also try this in his spare time. That will have consequences for the WiP of DataSeries for which @weefuzzy has another idea IIRC but for now it would be good to also get him to read it. Also if @AlexHarker wants to try this bespoke objects for his piece from that branch that might be a good idea. I will run the test on Monday as I am in compo/impro/code residency for the next days...

[tremblap]

it all works here. Is it possible that it doesn't bounce as much as before, or have I been lucky?

[tremblap]

I think I've been lucky. In all cases it works fine.

[jamesb93]

Working for me!