ggml : fix unnecessary f32 -> f16 -> f32 casts (mmla)

[ggerganov]

ref #4966

The struct block_q8_1 on the CPU uses float instead of ggml_fp16_t:

#define QK8_1 32
typedef struct {
    float d;               // delta
    float s;               // d * sum(qs[i])
    int8_t  qs[QK8_1];     // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");

[ggerganov]

@snadampal I haven't tested this change - please give it a try just in case

[snadampal]

Hi @ggerganov , LGTM. I have tested it on AWS Graviton3 based c7g instances.