V5 additions

[lorenss-m]

Note

Introduces remote BYOK support, tighter error propagation, stricter scenario validation, and broader HTTP client instrumentation.

• Remote BYOK support: Adds --byok CLI flag, byok config, and use_byok in SingleTaskRequest; forwarded via submit_rollouts while stripping api_key/model_client from remote agent_params.
• Agent error propagation: MCPAgent.run() now sets ctx.error on failures and step errors; new tests cover exception and step-level propagation.
• Scenario robustness: Environment.run_scenario_setup() now raises detailed errors for missing/invalid scenarios and malformed/empty content; MCPAgent.run() gives clearer messages when ctx.prompt is missing.
• Gateway auth rules: OpenAIChatAgent forbids custom api_key with HUD Gateway; CLI injects HUD API key for OpenAI-compatible when targeting gateway.
• HTTP instrumentation: Extends auto-instrumentation to aiohttp alongside httpx to inject trace/auth headers.
• Eval context stability: EvalContext.__aenter__ cleans up connections and contextvars on setup failure.
• Minor: log auto-respond errors; lazy-import questionary/typer; display byok in eval summary.

^{Written by Cursor Bugbot for commit 4ddaa33. This will update automatically on new commits. Configure here.}