Laser tensor PR rebase

[Vindaar]

Here it is finally. Sorry for taking so long.

This is a rebase of the laser tensor PR #420 onto the current state of arraymancer. The major complication was the restructure of the whole arraymancer library to support importing individual parts as submodules.

I had to fix a few things that came up:

• there was a self assignment line in pca.nim, which caused the tensor to be destroyed maybe? I removed it and now it works with arc. Relevant commit
  : 63b4705
  Maybe my =destroy implementation is at fault? @Clyybber maybe?
• I had to add a CpuStorageObj for the =destroy proc signature. That's the same as would happen implicitly I assume?
• there's still the issue with the supportsCopyMem not supported in type sections and the workaround:
  https://github.com/mratsim/Arraymancer/blob/laser-tensor-rebase/src/arraymancer/laser/tensor/datatypes.nim#L31-L36

As discussed on #420 this can cause issues where certain types (variant objects) match different branches in the type definition and in the code using when T.supportsCopyMem.
For that reason I will define a type class of types for which we know that they support copyMem which will then be used in the code as well as in the type definition to make sure laser based tensors are used for the vast majority of types used in practice.

I'll finish the latter maybe today and at the latest over the weekend. Wanted to open this PR first though, because I don't know how much more time I have today and to see what the CI says...

[Vindaar]

CI failure is because I forgot to add Complex to KnownSupportsCopyMem from here:
https://github.com/mratsim/Arraymancer/blob/laser-tensor-rebase/tests/tensor/test_filling_data.nim#L33

However, it's a bit worrying that this failed in the first place, indicating that non mem copyable types may have hidden bugs. It could be a good idea to run all basic tests in a mode where all types are treated as mem copyable? Maybe have a

when not defined(noPtrLenBackend):
type
  KnownSupportsCopyMem* = SomeNumber | char | Complex[float64] | Complex[float32]
else:
  KnownSupportsCopyMem* = distinct object

and then run tests with -d:noPtrLenBackend in addition?

edit:
I'm not sure if the test failure really is a failure or if the test is broken. copyFrom states that the two tensors should have the same shape:

  ## Copy the source tensor into the destination tensor.
  ## Both should have the same shape. If destination tensor is a view
  ## only the data exposed by the view is modified.

yet in the test they don't:

        let a = [[1,2],[3,4]].toTensor.reshape(2,2).astype(Complex[float64])
        var b = ones[Complex[float64]](4,1)

one is (2, 2), the other is (4, 1).

The thing is that the copyMem branch for contiguous tensors only performs a check on the size, but not the shape:

    if src.is_C_contiguous:
      assert dst.size == src.size
      omp_parallel_chunks(
            src.size, chunk_offset, chunk_size,
            OMP_MEMORY_BOUND_GRAIN_SIZE * 4):
        copyMem(
          dst.unsafe_raw_offset[chunk_offset].addr,
          src.unsafe_raw_offset[chunk_offset].unsafeAddr,
          chunk_size * sizeof(T)
        )

@mratsim: should copyMem perform an assert on the shape instead?
I'm going to fix the test case to use the same shape.

[Vindaar]

FYI: nim-lang/Nim#16185 is still an open issue, which makes the code here unsafe if used with ARC.

And a personal blocker: ggplotnim does not pass its dataframe tests (independent of ARC!). It throws an out of memory error. I'll try to debug that in the coming days. Once I figure out if it's a problem on ggplotnim's or arraymancer's side, I'd consider merging this PR. @mratsim still has to review it at that point though.