Feature/add phi4 convo

[EricApgar]

Mainly got the Phi4 model up to date with a similar use as the GPT model.

Closing out some issues that were technically closed by older PR's.

• Closes Restrict model specific libraries to specific installs #4
• Closes Add check to make sure model is loaded before using. #5
• Closes Add conversation structures. #7

[EricApgar]

Test suite added

Added a pytest test suite and GitHub Actions workflow as part of this PR.

New files:

• tests/test_gpt_oss_20b.py — loads the real GPT model and calls ask(), asserting a non-empty string response
• tests/test_phi4_multimodal_instruct.py — loads the real Phi4 model and calls ask() with a synthetic image, asserting a non-empty string response
• .github/workflows/tests.yml — triggers on PRs and pushes to main

Key design decisions:

• Tests call the real models rather than mocking, since LLM output consistency can't be tested — only that the model runs without errors
• Model cache path is passed via LLM_MODEL_CACHE as an inline environment variable (e.g. LLM_MODEL_CACHE=/path/to/cache pytest) so it doesn't persist in the shell
• Each model fixture uses scope='module' to load the model once per file, then tears down with del + torch.cuda.empty_cache() so both models can be tested sequentially without exceeding GPU memory
• Tests skip automatically if LLM_MODEL_CACHE is not set, so CI doesn't hard-fail without model weights
• The workflow is configured for a self-hosted runner (a machine with a GPU and local model weights). See Settings → Actions → Runners to register one, and add LLM_MODEL_CACHE as a repository secret.

To run locally:

uv sync --extra all
LLM_MODEL_CACHE=/home/yourname/Repos/model_cache pytest