trunk/835c64ac2dd146246f5433cd880ec24c0a4d0e21: [inductor] Fix block descriptor final shape expansion (#182936)

For fused pointwise & reduction kernels the pointwise operations need to be rank expanded to match [pointwise_numel, reduction_numel]. For block descriptor based load / store, currently this only applies if the number of dimensions matches the number of range trees minus the reduction trees. However this fails when there is discontiguity in the load / store.

For example:

range_trees = {x: ..., r0: ...}

block descriptor shape = [XBLOCK // 2, XBLOCK % 2]

num_reduction_dims = 1

so here the number of block descriptor dimensions won't be equal to (len(range_trees) - num_reduction_dims) (2 range trees, 1 reduction dim). and the determined final shape for this example will be [XBLOCK], instead of [XBLOCK, 1] (with the added reduction dim). This PR makes the check more robust by accounting for discontiguity.

Added a unit test for this fix. based on a compilation failure from GPT2 and Stable Diffusion XL torch compile run.

Pull Request resolved: #182936

Approved by: https://github.com/jansel