Tournament model args

[ianbulovic]

Adds support for passing a Model instance to the Tournament constructor instead of just a string. This allows the user to pass a model with custom model arguments. Also adds a new -M option to the tournament CLI for passing model arguments. For example with the CLI:

cnlp_llm tournament \
    path/to/data.tsv \
    path/to/prompt.prompt \
    --task ... \
    --pos-label ... \
    --model hf/meta-llama/Llama-3.1-8B-Instruct \
    --rounds 10 \
    --scheduler graph \
    -M "torch_dtype=auto" \
    -M "attn_implementation=flash_attention_2" \
    -M "trust_remote_code=True"

Or with the Tournament API:

from inspect_ai.model import get_model

model = get_model(
    "hf/meta-llama/Llama-3.1-8B-Instruct", 
    torch_dtype="auto", 
    attn_implementation="flash_attention_2",
    trust_remote_code=True,
)

tournament = Tournament(
    model=model,
    # ... other args
)

Resolves #41.

[wonjininfo]

This is awesome. I didn't know that I can do this. 👍 (parse_cli_args function)

[wonjininfo]

Looks good to me. Thanks a lot!!