feat: load multiple models

[vansangpfiev]

Issue link: #467

• use std::unordered_map to store all llama_server_context (lsc)
• refactor: move background thread to lsc, each lsc has its own background thread
• TODO: update documentation

We use mode_id as a key to find the model, so they have to be unique. That requires some changes in request parameter:
For completions:
Have to set value the model_alias for inferences/llamaCPP/loadmodel and inferences/llamaCPP/unloadmodel , this has to be the same as model parameter in inferences/llamaCPP/chat_completion
loadmodel/unloadmodel:

{
...
 "llama_model_path": "file_to_location"
  "model_alias": "model1"
...
}

chat_completion:

{
...
  "model": "model1"  
...
}

For embeddings:
The same for loadmodel/unloadmodel, for embedding request, we need to add model parameter
loadmodel/unloadmodel

{
...
    "llama_model_path": "e:/workspace/model/nomic-embed-text-v1.5.f16.gguf",
    "model_alias": "model1",
    "model_type": "embedding"
...
}

embedding

{
...  
  "input": "how are you",
  "model": "model1"
...
}

[louis-menlo]

The inference task queue is constructed with the model's parameters. It is shared among the model's operations, which is incorrect, right? Let's say it would be renewed every time a new model request with a larger n_parallel number comes in:

1. How will pending tasks be processed since it is renewed?
2. Will later requests change thread_num, causing side effects for previously loaded models?"

[vansangpfiev]

One solution (though it is not optimal) is that we can use a task queue for each model. Will change the code to make it works correctly first.

[louis-menlo]

Missing some words

[louis-menlo]

LGTM

[vansangpfiev]

Moved PR to janhq/cortex.llamacpp#14