Allow models loading on different backends

[alexpatcas]

Just like the title says. It would be great if we can load some models on vulkan and some models on rocm.
For example gpt-oss is great on vulkan, but gemma or any VL model struggles really bad.

Allowing different models on different back ends, would allo us the best of both words

[jeremyfowers]

This is possible today.

Load gpt-oss on ROCm:

curl -X POST http://localhost:8000/api/v1/load \
  -H "Content-Type: application/json" \
  -d '{
    "model_name": "gpt-oss-20b-mxfp4-GGUF",
    "llamacpp_backend": "rocm"
  }'

Load gemma on Vulkan:

curl -X POST http://localhost:8000/api/v1/load \
  -H "Content-Type: application/json" \
  -d '{
    "model_name": "Gemma-3-4b-it-GGUF",
    "llamacpp_backend": "vulkan"
  }'

Both can be loaded at the same time if you started the server with lemonade-server serve --max-loaded-models 2