Add Gemini 4 Argon and reasoning effort to Gemini Computer Use

[NikhilShahi]

Summary

Expose the API's new gemini-4-argon model and optional reasoning_effort for Gemini Computer Use (hyperbrowserai/browserctrl#1580). Gemini 3.8 Flash remains the default, and omitting reasoning_effort keeps each model's existing behavior.

Changes

• Add "gemini-4-argon" to GeminiComputerUseLlm.
• Add GeminiComputerUseReasoningEffort ("minimal" | "low" | "medium" | "high"), and add reasoning_effort to both the StartGeminiComputerUseTaskParams TypedDict and the legacy Pydantic model. It serializes as reasoningEffort. The API rejects "minimal" for Argon, 3.8 Flash, and 3.7 Flash.
• Re-export GeminiComputerUseLlm and GeminiComputerUseReasoningEffort from hyperbrowser.models, matching the other agents' Llm and effort types.
• Add request-compat tests (mapping and legacy produce the same payload, and an omitted effort is not sent) and mypy/Pyright fixtures, including an invalid "xhigh" effort.
• Bump the version to 1.12.1.

client.agents.gemini_computer_use.start_and_wait(
    {
        "task": "Go to Hacker News and tell me the title of the top post",
        "llm": "gemini-4-argon",
        "reasoning_effort": "high",
    }
)

Validation

• ruff check and ruff format --check on the changed files. tests/typecheck/invalid_requests.py already had an unrelated formatting issue before this change.
• pytest tests --ignore=tests/sandbox/e2e: 814 passed, including the mypy/Pyright typing-contract tests.