VORTEXM4: SYRK/SYR2K diagonal blocks are expensive with GEMM_UNROLL_MN 64 #6148

On VORTEXM4 (GEMM_UNROLL_M 64 since #6074), the generic SYRK/SYR2K driver spends much time on the diagonal blocks. driver/level3/syrk_kernel.c computes each GEMM_UNROLL_MN x GEMM_UNROLL_MN diagonal block (64 x 64 here) into a temporary buffer zeroed with GEMM_BETA and then adds one triangle into C. A sampling profile of SSYRK at n = 256, one thread, on an M4 Pro shows ssyrk_kernel_L plus sgemm_beta zero-fill at about 20% of the time, against under 10% for the SME2 compute (the rest is packing, see #6146).

Measured against Accelerate on the same machine (one thread): SSYRK 0.18x / 0.36x / 0.73x at n = 64 / 256 / 1024, 0.75-0.93x at 4096. trans = T is slower than trans = N (1074 vs 1370 GFLOPS at n = 1024).

Options:

• compute the diagonal block directly into C with a masked store (the SME2 kernels can predicate rows and columns), which needs a per-target hook in the driver;
• a smaller diagonal block for targets with a large GEMM_UNROLL_MN;
• check why the trans = T path is slower (packing of A^T with sgemm_otcopy).