Fix LLM GIF parsing robustness and add CLI producer

[apshoemaker]

Motivation

• Real-world LLM responses often include surrounding prose or fenced code blocks which caused a raw json.loads to fail with JSONDecodeError; the project needs a producer that tolerates those formats and emits ND75-ready GIF bytes.
• Provide a reusable, test-covered producer CLI to generate GIFs from natural-language prompts so users can pipe LLM-generated animations into the existing push workflow.

Description

• Add nd75_screen/llm_gif.py which extracts text from mixed Anthropic response blocks, scans and decodes embedded/fenced JSON payloads, validates storyboard shape, renders ND75-sized frames with PIL, and serializes them to an animated GIF via frames_to_gif.
• Add nd75_screen/cli/llm.py CLI (nd75-llm-gif) that creates an Anthropic client from ANTHROPIC_API_KEY, accepts a prompt and --model override, and writes generated GIF bytes to stdout.buffer for pipe composition.
• Add tests tests/test_llm_gif.py and tests/test_cli_llm.py covering fenced JSON parsing, ignoring non-text blocks, clear ValueError on missing JSON payloads, CLI stdout emission, and model override behavior.
• Update packaging and docs by adding anthropic>=0.84.0 and a nd75-llm-gif script entry in pyproject.toml, and adding usage/examples for the new CLI to README.md.

Testing

• Ran uv run pytest tests/test_llm_gif.py tests/test_cli_llm.py -q and the new tests passed.
• Ran uv run pytest tests/ -q and the full test suite passed (81 passed).
• Verified the CLI help with uv run python -m nd75_screen.cli.llm --help which printed the expected usage information.

---

Codex Task