Leverage the existing GGML_F32_VEC helpers to vectorize ggml_vec_set_f32 for faster fills

[sirus20x6]

Add a SIMD path to ggml_vec_set_f32, broadcasting the fill value with the existing GGML_F32_VEC helpers

Keep the scalar tail for leftover elements and non-SIMD builds

[sirus20x6]

microbenchmarks show sometimes very little change, but sometimes a nice bump

Baseline (pre-SIMD helpers)

add1
  n=128     throughput=90.09 GB/s
  n=1024    throughput=117.64 GB/s
  n=8192    throughput=55.15 GB/s
  n=65536   throughput=57.87 GB/s
  n=524288  throughput=51.29 GB/s
acc
  n=128     throughput=78.97 GB/s
  n=1024    throughput=118.10 GB/s
  n=8192    throughput=57.20 GB/s
  n=65536   throughput=57.96 GB/s
  n=524288  throughput=48.29 GB/s
acc1
  n=128     throughput=88.28 GB/s
  n=1024    throughput=114.89 GB/s
  n=8192    throughput=131.41 GB/s
  n=65536   throughput=114.31 GB/s
  n=524288  throughput=87.01 GB/s
mul
  n=128     throughput=87.03 GB/s
  n=1024    throughput=61.81 GB/s
  n=8192    throughput=40.49 GB/s
  n=65536   throughput=34.54 GB/s
  n=524288  throughput=31.66 GB/s

Current branch (with SIMD helpers)

add1
  n=128     throughput=100.72 GB/s
  n=1024    throughput=141.98 GB/s
  n=8192    throughput=55.42 GB/s
  n=65536   throughput=59.20 GB/s
  n=524288  throughput=51.91 GB/s
acc
  n=128     throughput=80.21 GB/s
  n=1024    throughput=134.74 GB/s
  n=8192    throughput=68.63 GB/s
  n=65536   throughput=56.20 GB/s
  n=524288  throughput=48.49 GB/s
acc1
  n=128     throughput=89.30 GB/s
  n=1024    throughput=142.30 GB/s
  n=8192    throughput=142.24 GB/s
  n=65536   throughput=118.68 GB/s
  n=524288  throughput=90.02 GB/s
mul
  n=128     throughput=86.29 GB/s
  n=1024    throughput=95.78 GB/s
  n=8192    throughput=42.58 GB/s
  n=65536   throughput=32.26 GB/s
  n=524288  throughput=31.22 GB/s

[ggerganov]

We should make the code consistent about how it handles the leftovers. Here we duplicate the scalar code, while in ggml_vec_add_f32 above we use a common loop iterator. I think we should do the same as in ggml_vec_add_f32.

[sirus20x6]

sure thing. latest brings simd/scalar functions inline with each other.