Adding TRT-LLM config key

[aspctu]

🚀 What

This PR adds a new key to the config.yaml: trtllm. We use this for models that need engines built or to serve engines. The spec is:

trtllm:
  engine_repository: "<path_to_trt_engine_repository>"
  tokenizer_repository: "<path_to_tokenizer_repository>"
  tensor_parallel_count: int
  pipeline_parallel_count: int
  build_configuration: # Used to build engines
    huggingface_ckpt_repository: "<path_to_huggingface_checkpoint_repository>"
    base_model_architecture: "llama"  # Can be "llama" or "mistral"
    max_input_len: int
    max_output_len: int
    max_batch_size: int
    quantization_type: "no_quant"  # Can be "no_quant", "weights_only", "weights_kv_int8", "smooth_quant"
    gather_all_token_logits: bool
    multi_block_mode: bool
    calibrate_kv_cache: bool

This PR also updates pydantic to a newer version.

💻 How

🔬 Testing

Integration tests passed for pydantic change. Also tested the new key with various configs locally via manual context builds.

[bolasim]

lgtm overall but would love an explanation of why we have to write the serialization manually. I thought we could rely on pydantic for that

[bolasim]

also, I think this PR is missing changes to bump the pydantic version on the server. I thikn that's probably okay assuming all the integration tests pass.

[bolasim]

I left two suggestions that let you drop all the to/from dict functions. lemme know if it works.

[joostinyi]

@aspctu can you update the example from the description for the most recent interface?