mul of two strided vectors fails with NPE

[katox]

(def a (tensor [3 2 1] :float :ncw))
;=> #'user/a
(desc a)
;=> #uncomplicate.diamond.tensor.TensorDescriptorImpl {:shape [3 2 1], :data-type :float, :layout [2 1 1]}
(.stride a)
;=> 1
(.stride (view-tz a [3 1 1]))
;Execution error (NullPointerException) at uncomplicate.diamond.internal.dnnl.tensor.DnnlTensor/stride (tensor.clj:442).
;Cannot invoke "uncomplicate.neanderthal.internal.api.Block.stride()" because "this.vector_view" is null

This is a simplified case. In real usage this codepath is executed when foldmap is to be internally used on strided tensors (non-contiguous vctr-view).

For instance (mul strided-tz-view strided-tz-view)

[katox]

Example:

(mul a a)
=> {:shape [3 2 1], :data-type :float, :layout [2 1 1]} (0.0 0.0 0.0 0.0 0.0 0.0)
(mul (view-tz a [3 1 1]) (view-tz a [3 1 1]))
Execution error (NullPointerException) at uncomplicate.diamond.internal.dnnl.tensor.DnnlTensor/stride (tensor.clj:442).
Cannot invoke "uncomplicate.neanderthal.internal.api.Block.stride()" because "this.vector_view" is null

[blueberry]

Elementwise product is not supported by non-contiguous DNNL tensors. You can use Elementwise multiplication instead (please see dnnl-binary-alg-kind for the list of supported ops).

[katox]

Yes, but usually there is a check instead of NPE.

[blueberry]

I implemented it that way intentionally (the code was simpler), but let's leave this open, so I introduce a check-and-throw macro for this case.