Inconsistencies in model paths on Hugging Face

[msluszniak]

As mentioned in #382 I'm moving this into separate issue:
Inconsistencies in naming across hugging face repository:

• Only for llama3_2-3B_qat_lora.pte there is - between name and number of parameters, while for the rest is _.
• For WHISPER_TINY_DECODER en is concatenated in the first part of URL with . and not with - as in the rest of examples.
• There is word all in efficientnet_v2_s_coreml_all and only for that model.
• For some of the models initial name is concatenated with - and the model name as well - phi-4-mini and phi-4-mini_bf16.pte while in others there is - and _ e.g. qwen-2.5 and qwen2_5_3b_8da4w.pte.
• Some model names adds backend info at the beginning: xnnpack_whisper_encoder.pte and some on the end or almost end: style_transfer_udnie_xnnpack.pte.
• ssdlite320-mobilenet-v3-large v3 here is concatenated with - but in model name not: ssdlite320-mobilenetv3-large.pte.
• For craft detector, number means size of the input tensor while for the other suffixed numbers means model sizes.
• Some models B and M meaning billion and million of parameters is capitalized in both first part of url and name: smolLm-2-135M and smolLm2_135M_bf16.pte and for some not: qwen-2.5-1.5B and qwen2_5_0_5b_8da4w.pte.

This should be unified in a separate task

[msluszniak]

Ok, we agreed that models should be named with model name & precision and be in correct directory according to backend

[NorbertKlockiewicz]

Right now there's a total mess on our hugging face repositories. There's no convention which would help us unify those, that's why I propose such rules:

• the pte file name: {modelName}_{size}_{backend}_{precision}.pte (e.g yolo26_m_coreml_int8.pte)
• a directory for each backend, when entering the repo in root there should be a directory for each backend to which model was exported, a config.json, README.md and if applies tokenizers
• we should have at most two versions of model for each backend (one in original precision - based on backend support fp16 or fp32 and one quantized which should be a default
• the config.json should have an information about the methods and their input + outputs shape and types