Tweak KleidiAI's FP16 matmul algorithm (pytorch#4416)

Summary:
Pull Request resolved: pytorch#4416

X-link: facebookresearch/FBGEMM#1488

Hoist memory loads from the outer loop

Intention is to prevent these loads from displacing cache lines, as they may contain matrix data.
Similarly, the loads are likely to inccur in cache misses after the first iteration. Executing the inner loop will probably fill the cache with matrix data.

Benchmarks repeatedly show a throughput improvement of around 1%.

before:
P1854747253

after:
P1854747141

Reviewed By: YifanYuan3

Differential Revision: D77459967

fbshipit-source-id: 01eb4fc004ba055823551d843f7bf7728caa74a8

Author: Nicoshev