Wrong Gemm library translation on sliced matrices

[luigifusco]

Describe the bug The library expansion for the Gemm operation using cublas is wrong on sliced matrices. This likely applies to other library expansions. Correct behavior would require the called library function to support strided matrices, which is not the case. The current implementation will detect the correct number of rows/columns involved as well as the row offset, but not the column offset.

The "pure" implementation works correctly.

To Reproduce bug.zip

Expected behavior Translation should fail or fallback to a pure implementation with a warning