fix(openai): cache stable prompts explicitly

[solvemproblr]

Issue

GPT-5.6 implicit prompt caching writes through changing user input, so independent rollouts cannot reuse a shared system prompt and may pay for cache writes that never hit.

Solution

• Put the stable system prompt in a developer content block with an explicit cache breakpoint for GPT-5.6 and later.
• Use explicit-only caching so changing rollout input remains outside the cached prefix.
• Expose prompt_cache_key for high-volume request grouping and retain cache_write_tokens in normalized usage.
• Preserve the existing request shape for earlier GPT models.

Validation

• uv run --extra modal pytest -q — 1116 passed
• uv run ruff check .
• uv run ruff format . --check
• uv run ty check hud/agents/openai/agent.py hud/agents/types.py hud/agents/tests/test_openai_agent.py