feat: load scaled FP8 weights without upfront conversion

[leejet]

Summary

• preserve E4M3 and E5M2 safetensors weights in their original one-byte representation
• cast active FP8 linear weights to BF16 at runtime and apply scalar or per-channel weight scales
• normalize .scale_weight tensor names to .weight_scale
• align sd_type_t with the new ggml FP8 types and update the ggml submodule

Related Issue / Discussion

N/A

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.