Add while loop section to parallel model

[jdahm]

Describes the syntax and limitation of while loops.

[havogt]

Thanks! This summarizes what I was missing to review the implementation PR.

I was thinking a bit more about it and came to the conclusion, that the restriction actually doesn't apply, because it is already covered by rule 4. Therefore it would be good to actually add proper checking for restrictions described here (which probably are currently missing) and remove the restriction on the while loop in the implementation.
Please double check if I got it right or if I missed an important that makes it special.

[havogt]

Suggested change

	however due to the blocking model used, there is no way to enforce
	synchronizations between the nested statements, or the mask update.
	This is a subtle but important point to remember when employing
	while loops. The final parallel model behaves as
	```python
	for k_ in range(k, K):
	parfor ij:
	mask[i, j] = (a[i, j, k_] < b[i, j, k_])
	while any(mask):
	parfor ij:
	if mask[i, j]:
	c[i, j, k_] += 1
	a[i, j, k_] += 1
	mask[i, j] = (a[i, j, k_] < b[i, j, k_])
	```
	The conclusion from this is that the user is not allowed to write to
	variables used with a horizontal offset either in the mask or elsewhere in
	the body of the while. The gtscript frontend implemented in gt4py contains checks
	to ensure that a user cannot write code incompatible with this restriction.

[jdahm]

within a computation, it is illegal to write to an external (non-temporary) field (or aliases pointing to the same memory location) if it is also read with horizontal offset; this rule does not apply to temporaries,

True, but I believe there still could be issues with the update of the mask. Consider this:

a = input_field # a is a temporary field
while a[1, 0, 0] < 0:
    a += 1
    ...

I suppose the rule should be that there are no horizontal offsets (or vertical offsets for that matter) in the while condition. Does that sound right to you?

[havogt]

I think it makes sense. It would require a temporary with halo of the size of max iterations that the while loop could have (which is unknown), right?

[havogt]

The important point is, if I am not mistaken, that the restriction doesn't apply to temporaries and therefore doesn't apply to the mask. Does it make sense?

[havogt]

Looks good.