Optimize bli_gemm_haswell_asm_d6x8: reduce latency & improve throughput (#895)

Details:
- Reorder instructions to reduce pipeline stalls and dependencies
- Adjust register usage to improve data reuse
- Tweak prefetching so cache warming is more effective
- Since gemm kernels now support general strides, corrected macro to
  enable optimized assembly kernel for general stride inputs

Author: harsdave