fix(init): clean up agent onboarding leftovers and persist onboarding source

[junhsss]

Addresses items 1–3 of #91 . Item 4 (login's 5-minute localhost callback wait) is covered by the device flow work in #92 and is intentionally left out here.

1. Remove orphaned init_agent_guide.md

steel init --agent stopped printing this guide when --agent was repurposed to mean "auto-accept prompts". The file has had zero references since, and its content is stale (predates the steel-dev/skills install flow and claims cross-device auth works, which the 127.0.0.1 callback does not allow).

2. Sync --agent docs with actual behavior

• install.sh docstring and the comment above the steel init --agent invocation still described the removed print-and-exit behavior.
• The Getting Started block in steel --help said the same thing.

Both now describe what --agent does: run non-interactively, auto-accept prompts, print agent-friendly output.

3. Persist onboarding attribution and attach it to telemetry

install.sh --from <agent> exported STEEL_ONBOARDING_FROM, but the only consumer was a status!() line in init. login_completed fired with empty properties, and since login usually happens in a later shell, the env var was gone by then anyway.

Now:

• steel init persists the value to config.json as onboarding.source.
• Telemetry loads onboarding.source at startup and attaches it as onboarding_source on every event, so login_completed and first-session events carry it even in a later shell.
• Within the same steel init process, the source is pushed into the telemetry state immediately so login_completed picks it up before the config is re-read.