Issue #2533: Preserving the buffer type in linear algebra

[kmp5VT]

An attempt to fix buffer type linear algebra issue. It looks like the CuArray function is called many times and there is no book keeping of the memory_type. I manually added to the CuArray function calls. One could also make this part of the function call. See Issue #2533

[maleadt]

Thanks for the PR! It is often easier to use the similar function, which can take an eltype and dims:

similar(array, [element_type=eltype(array)], [dims=size(array)])

Create an uninitialized mutable array with the given element type and size, based upon the given source array. The second and third arguments are both optional, defaulting to the given array's eltype and size.
The dimensions may be specified either as a single tuple argument or as a series of integer arguments.

[kmp5VT]

@maleadt Oh yes, similar is a better way to go. I will update these changes with similar, thanks!

[maleadt]

Yeah looks better with similar 🙂 The memory_type now can go, though?

[maleadt]

LGTM, CI failure unrelated.

Why the partial revert though?

[kmp5VT]

@maleadt I wasn't sure what the failures were related to. With the changes I reverted, I haven't yet had time to verify that they work as expected so I wasn't sure if those changes were causing the check failures. I have more time today and can check that function and re-push the changes.

[kmp5VT]

@maleadt I was able to check that the gebrd! code does work properly if all the tensors are on unified memory. I added back those changes. Thanks!