fix(pi): use model-specific GPT thinking levels

[dgokeeffe]

Summary

• declare model-specific GPT reasoning controls in Pi's custom OpenAI Responses entries
• add complete GPT-6 Astra metadata: text/image input, 1,050,000-token context, and 128,000 max output
• map Pi's off control to the Responses API's none value only where supported
• suppress unsupported minimal, low, and off choices for newer and Pro variants
• add regression coverage for Astra, GPT-5.4, GPT-5.5, GPT-5.6, Pro, and legacy GPT-5 models

Fixes #505

Model mappings

• GPT-6 Astra: low, medium, high, xhigh, max
• GPT-5.6: none, low, medium, high, xhigh, max
• GPT-5.5 and GPT-5.4: none, low, medium, high, xhigh
• GPT-5.4/5.5 Pro: medium, high, xhigh
• legacy GPT-5: retain the existing compatible minimal, low, medium, high behaviour

Validation

• uv run pytest tests/test_agent_pi.py — 52 passed
• uv run ruff check src/ucode/agents/pi.py tests/test_agent_pi.py
• uv run ruff format --check src/ tests/
• git diff --check

The full suite ran with 2,177 passed and 37 skipped. Four unrelated environment-sensitive tests failed in Claude smart-routing PTY/socket and Claude user-agent e2e coverage; the Pi suite and formatting/lint checks passed.