Remove hard agent-turn timeout from cron/task schedulers

[LIU9293]

Why

The cron scheduler (packages/core/cron/scheduler.ts) and one-time task scheduler (packages/core/tasks/scheduler.ts) both wrapped agent.sendMessage in a 2h hard timeout via withTimeout (configurable via ODE_CRON_AGENT_TIMEOUT_MS / ODE_TASK_AGENT_TIMEOUT_MS).

This bound was tripping on legitimate long-running agent workflows — Sentry triage, Linear board grooming, multi-PR sweeps, scheduled audits — and surfacing in Slack as:

Cron job failed: <title>
Request timed out

even though the underlying agent was making progress. Verified on this machine: two recent runs of `Sentry检查` and `PM推进` both failed with `CronStepTimeoutError: Cron agent turn timed out after 7200000ms` (= exactly the 2h default).

What

• Drop the `withTimeout` wrap around `agent.sendMessage` in both schedulers.
• Remove the `CRON_AGENT_TIMEOUT_MS` / `TASK_AGENT_TIMEOUT_MS` constants and the corresponding `ODE_CRON_AGENT_TIMEOUT_MS` / `ODE_TASK_AGENT_TIMEOUT_MS` env vars.
• Leave a comment explaining the decision and the remaining safety nets.

What we keep

• `CRON_PREPARE_TIMEOUT_MS` / `TASK_PREPARE_TIMEOUT_MS` (2 min): still wraps session creation + worktree setup so a wedged `getOrCreateSession` / `git worktree add` can't hold the in-process `runningJobIds` / `runningTaskIds` lock indefinitely (which would otherwise make manual "Run now" return 409 forever).
• `reconcileInterruptedCronJobs` / `reconcileInterruptedTasks`: still recovers stuck runs on daemon restart.
• The agent adapter's own per-request error handling.

Verification

• `bun run typecheck` clean.
• `bun test packages/core/test` → 78 pass / 1 skip / 0 fail.