feat(chat): add friendly name aliases for chat templates (fixes #976)

[octo-patch]

Summary

Closes #976.

Users currently cannot pass friendly names like "llama2" or "ChatML" to the chat_template parameter — they must supply the full Jinja2 template string. This PR implements the alias system that was explicitly called out in two TODO comments in chat.py:

TODO [HN]: Add an alias system so we can instantiate with other simple keys (e.g. "llama2" instead of the full template string)

Changes

• guidance/chat.py — updates load_template_class() to check a new CHAT_TEMPLATE_NAME_MAP dictionary before falling through to the Jinja2 template string cache. Lookup is case-insensitive and strips spaces, dashes, underscores, and dots, so all of these work:

  model = Transformers("meta-llama/Llama-2-7b-chat-hf", chat_template="llama2")
  model = Transformers("meta-llama/Llama-2-7b-chat-hf", chat_template="Llama2")
  model = Transformers("meta-llama/Llama-2-7b-chat-hf", chat_template="llama-2")
  model = Transformers("meta-llama/Llama-3.2-1B-Instruct", chat_template="Llama 3.2")
  model = Transformers("Qwen/Qwen3-8B", chat_template="qwen3")

• CHAT_TEMPLATE_NAME_MAP — a new module-level dict mapping normalized alias keys to template classes. Covers all templates currently defined in the module: ChatML, Llama 2/3/3.2, Phi-3 mini/small/medium, Phi-4 mini, Mistral 7B Instruct, Gemma 2 9B, Qwen 2.5, and Qwen 3.

• tests/unit/test_chat.py (new) — unit tests covering all alias variants, case-insensitivity, separator stripping, None fallback, unknown-name warning, and map-key normalization invariant.

Backwards compatibility

Existing code that passes full Jinja2 template strings continues to work unchanged — the name map check only runs first; if no alias matches, the code falls through to CHAT_TEMPLATE_CACHE exactly as before.