Absolute K indexing fix: memlet volumn when mixed with cartesian

[FlorianDeconinck]

As per inline comment:

Today, we are looking at cases where absolute read and cartesian writes live in the same vertical loop e.g.

  with computation(PARALLEL), interval(...):
      _ = field.abs(K=x)                      # noqa: ERA001
      ...
      field = a + b                           # noqa: ERA001

This code is valid, the issue with DaCe bridge is that the volume given to the memlet differs in this case:

• read memlet (absolute indexing): full array, we don't know
• write memlet (cartesian): 1 - we know exactly
  Since those apply to the same underlying array, DaCe is not too happy about it. The workaround is to use the "variable K offset" pipeline for the cartesian to match the wider volume and at the same time keep the relative index correct.

[romanc]

I think this is fine as a workaround for this branch. If there's an easy test-case, we could take that and add it to the (mainstream) gt4py/cartesian testcases. That would make sure of "it'll go away (tm) in the new bridge".

[FlorianDeconinck]

I have the goods, I'll put in a test case.