metal : improve F32, F16 and BF16 mat-vec multiplication

[ggerganov]

Better threadgroup utilization for the floating-point mat-vec kernels:

Model	Test	t/s master	t/s gg/metal-mul-mv-opt-2	Speedup
qwen3 1.7B BF16	tg32	127.14	141.96	1.12
qwen3 1.7B F16	tg32	127.44	142.26	1.12
qwen3 1.7B all F32	tg32	83.96	86.97	1.04

Also some TG gains for MoE models of any quantization since they use F32 matrix multiplication in the FFN:

Model	Test	t/s master	t/s gg/metal-mul-mv-opt-2	Speedup
gpt-oss 20B MXFP4 MoE	tg32	128.57	132.93	1.03
qwen3moe 30B.A3B Q4_0	tg32	100.39	102.86	1.02