This adds an optimization to save rows/cols of sub-matrices stored in SLM into registers prior to summing over delta_m, delta_n and updating private result with products.

This does not make any difference in performance of the compiled code though due to compiler optimization being effective as loops are unrolled, but makes compiler's job easier and the intent clearer.

• I verified that compiler output for the kernel with this change is identical to the output before it (checked on Intel Max GPU).
• I checked that code compiles for CUDA backend and test suite run is green

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• If this PR is a work in progress, are you opening the PR as a draft?