docs(ollama): Ollama launcher + setup docs + fix README misdirection (closes #118)

[MoonlightByte]

Summary

• Fixes README.md lines 458-459 which misdirected users to configure a non-existent endpoint setting in config.py (issue [DOCUMENTATION] Ollama support #118).
• Adds a dedicated Ollama integration section in README.md mirroring the existing LM Studio section.
• Adds run_with_ollama_direct.bat and run_with_ollama_direct.sh — self-contained launchers that handle Ollama alias setup inline on first run. Both of the game's internal model tiers are aliased to a single user-chosen Ollama model (matches LM Studio's single-model behavior; avoids VRAM thrashing).
• Adds OLLAMA_SETUP.md and OLLAMA_QUICKSTART.txt.

UX parity with LM Studio

Step	LM Studio	Ollama
1	Install LM Studio	Install Ollama
2	Load a model in the LM Studio UI	`ollama pull `
3	Double-click `run_with_lmstudio_direct.bat`	Double-click `run_with_ollama_direct.bat`

On first run, the launcher auto-creates two aliases (`gpt-4.1-2025-04-14` and `gpt-4.1-mini-2025-04-14`) both pointing at the user's single pulled model. No separate setup script, no manual `ollama cp` calls. Setting `OLLAMA_MODEL=` before launch forces re-aliasing to a different model.

Out of scope

• No Python code changes. The `multi-model-refactor` branch is the long-term home for first-class Ollama integration.
• Image generation and TTS remain broken under Ollama (no compatible endpoint upstream). Documented in the Known Issues section.
• `README.md` lines 416 and 954 have the same `config.py` misdirection class-bug; left for a follow-up docs sweep because they don't block issue [DOCUMENTATION] Ollama support #118.
• No `run_with_lmstudio_direct.sh` added for macOS/Linux — that gap predates this PR.

Test plan

End-to-end testing with a live Ollama server was not performed on the author's machine (local Ollama is 0.1.29, below the documented 0.4+ floor, and the daemon wasn't running). Static verification was done instead:

• Shell script `bash -n` clean
• Spec compliance review: files match plan verbatim
• Code quality review approved after 3 fix commits (macOS bash 3.2 portability, `OLLAMA_MODEL` override working correctly, `curl` timeout, `findstr /L /C:` literal match)
• No Python files modified
• LM Studio integration section preserved; no LM Studio files touched
• First-run with one pulled model auto-creates both aliases and starts the game (requires Ollama 0.4+ to test)
• Re-run with aliases present skips setup and starts immediately
• With multiple pulled models, launcher aborts and requires `OLLAMA_MODEL`
• With `OLLAMA_MODEL` set, launcher re-points aliases even when they exist
• A game action round-trips through Ollama on `/v1/chat/completions`
• `run_with_lmstudio_direct.bat` still works (Windows regression check)

Closes #118.