Add gpt-5.4 models and request options

[sderev]

What changed

• Add lmterminal/model_registry.py and move model aliases, pricing data, and tokenizer fallbacks into it.
• Add GPT-5.3, GPT-5.4, GPT-5 Pro, o3-pro, and newer Codex model names to the CLI.
• Add --reasoning-effort and repeatable -o/--option key=value passthrough for Chat Completions requests.
• Apply gpt-5.4 short vs long prompt pricing in --tokens.
• Update README, template comments, changelog fragment, and tests.

Why

• OpenAI added new model names and GPT-5.4 long-context pricing.
• Users need a first-class reasoning control and a generic escape hatch for newer Chat Completions options.
• One shared registry reduces drift between model validation and prompt-cost estimation.

How to test

• gate
• uv run pytest tests/cli_test.py tests/gpt_integration_test.py tests/lib_test.py
• Run uv run lmt models and confirm the new GPT-5 / Codex / o3-pro names appear.
• Run uv run lmt --tokens -m gpt-5.4 "hello" and confirm the prompt cost output includes a pricing tier line.
• With a valid API key, run uv run lmt -m gpt-5.4 --reasoning-effort high -o verbosity=low "hello" and confirm the request succeeds.
• Run uv run lmt -o temperature=0.9 "hello" and uv run lmt -o n=2 "hello" and confirm the CLI rejects those -o overrides.

Risk/comp notes

• -o/--option uses light parsing and relies on the OpenAI API to reject invalid model-specific combinations.
• Hidden legacy price-only models remain supported by the shared pricing helpers and template-driven usage, but they are not exposed by lmt models or accepted via -m.

Changelog fragment: yes