feat: support OpenAI-compatible local providers

[aladin7]

Adds an opt-in OpenAI-compatible local LLM provider while preserving Ollama as the default.

Fixes #16.

What changed

• Adds an assistant-provider selector for Ollama and OpenAI-compatible local servers.
• Adds /v1/models discovery, /v1/chat/completions requests, optional bearer-token support, and OpenAI tool-call argument normalization.
• Keeps separate persisted URL/model values for each provider so switching does not overwrite the other configuration.
• Generalizes health-check errors and settings labels in English, Italian, and German.
• Documents llama.cpp setup and the provider-specific configuration.
• Adds provider adapter, model discovery, settings persistence, and backward-compatibility tests.

Validation

• python -m py_compile assistant.py config.py main.py settings_window.py locales.py test_assistant.py test_settings_provider.py
• python -m unittest discover -v — 54 tests pass
• ruff check test_assistant.py test_settings_provider.py
• Real loopback HTTP contract test verified GET /v1/models, POST /v1/chat/completions, tool schema submission, and returned tool-call normalization.
• Real llama.cpp b9982 inference with Qwen3-1.7B Q4_K_M correctly selected and populated WritHer's list_notes, save_note, and create_appointment tools; the full assistant.process() path also passed.
• Tk settings-window smoke test verified the provider, model, and URL controls render and synchronize correctly.

An end-to-end manual run was also completed successfully: WritHer captured microphone input, transcribed it, sent the command to the local llama.cpp server, parsed the returned tool call, and executed the requested action.

AI assistance: OpenAI Codex was used for specification research, implementation, and local validation.