Refactor: Move matrix packing outside GEMM kernels

In class tinyBLAS_PPC, previously, packing of input matrices A and B
was performed on-the-fly within each GEMM microkernel.
This patch refactors the code to decouple packing from kernel
by introducing a preprocessing step that packs matrices once before any kernel is invoked.

Benefits:
- Enables better memory locality and data reuse
- Simplifies the kernel logic by focusing purely on computation
- Improves overall GEMM performance, especially for large matrix sizes

Signed-off-by: Shalini Salomi Bodapati <[email protected]>

Author: shalinib-ibm