Canonical dtype casting for GGUF conversion: {F32, F16, BF16} sources to any
writable GGMLType, including Q8_0 block quantization. Bit-for-bit the same
math as the per-model converters (which predate this crate and keep their own
verified copies).
Read any supported checkpoint (including an existing GGUF) and rewrite it at a
different dtype, carrying metadata through. Shared by the generic zsfm convert
command and by every per-model CLI’s cached-GGUF fast path: once a repo has been
downloaded and converted to a canonical F32 GGUF once, later requests for a
different dtype recast from that cache instead of re-downloading from HuggingFace.