metal : fix kernel_norm (fixes Falcon on Metal) (#3057)
* metal : fix kernel_norm
ggml-ci
* metal : put warning in kernel_norm to not combine the loops
* metal : restore original F16 mat-vec multiplication
It works after the norm fixes
* common : don't do warm-up with more than n_batch tokens (close #3058)
ggml-ci
* metal : minor
