[codex] Improve LiteLLM OpenAI Gemini dispatch

[Tibo2403]

What changed

• Expanded the existing Codex LiteLLM router with task-oriented aliases: codex-light, codex-default, codex-long, and codex-deep.
• Added weighted OpenAI/Gemini model groups, cooldowns, context-window fallbacks, pre-call checks, and health-check routing to litellm-cost-routing.yaml.
• Updated the Python router to classify long-context work, support a gemini provider preference, and route complex work to codex-deep.
• Updated the PowerShell launcher to request GEMINI_API_KEY as an optional in-memory secret and clear it on shutdown.
• Refreshed docs and tests for the new dispatching behavior.

Security

• API keys remain environment-variable only: no provider secrets are committed.
• Ran the local high-confidence secret scan before pushing.

Validation

• python -m unittest scripts.python.tests.test_codex_cost_router -> 18 tests passed.
• LiteLLM YAML parsed successfully and verified simple-shuffle plus context fallbacks.
• Secret scan: no high-confidence secret patterns found.

[Tibo2403]

Added the optional local web key session flow in dfe5cca.

What it adds:

• scripts/python/codex_key_session_web.py: localhost-only page for entering OpenAI/Gemini/HF keys for one work session.
• scripts/python/Start-CodexKeySessionWeb.ps1: starts the page on 127.0.0.1:8787.
• scripts/python/Install-CodexLocalLiteLLMAssets.ps1: copies the local assets to ~/.codex/litellm-proxy and installs a config.yaml without master_key for Codex Desktop compatibility.

Security note: keys are passed only to the LiteLLM subprocess environment and are not written to disk; request logging is suppressed.

Validation rerun:

• python -m unittest scripts.python.tests.test_codex_cost_router scripts.python.tests.test_codex_key_session_web -> 21 tests passed.
• YAML validation OK.
• Secret scan OK.