ml: detect model parameters at conversion time and write the config to gguf

* introduce `model_file` to read key-value data from gguf files
* conditionally set `cwhn` flag based on gguf tensor data layout