Make the agent step timeout configurable and record spend of failed runs

[iyernaveenr]

Problem

Any agent turn longer than thirty minutes is aborted with Async step timed out after 1800s, regardless of whether it was making steady progress (observed with a long single-agent writing task that was still producing output when the ceiling hit). The limit is hardcoded, so deployments running long-form tasks cannot adjust it. In addition, a run that fails this way leaves its history row with zero tokens forever, hiding the real spend of the aborted work.

Root cause

• ListenChatAgent hardcodes step_timeout: float | None = 1800; the underlying agent wraps each non-streaming step in asyncio.wait_for with that value, so the whole turn is cancelled at the ceiling. The screenshot toolkit repeats the same hardcoded fallback.
• On the frontend, the error step handler marks the task failed but never updates the run's history row, so tokens stays at its initial zero even though per-exchange token counts had been accumulated client-side.

Fix

• Backend: read the default step timeout from AGENT_STEP_TIMEOUT_SECONDS (seconds; same 1800 default when unset; a non-positive value disables the per-step ceiling; invalid values log a warning and keep the default). The screenshot toolkit fallback uses the same source instead of a second hardcoded constant. Explicit step_timeout arguments (e.g. the short validation timeouts) behave as before.
• Frontend: when a run ends in an error, record the tokens accumulated so far into the run's history row (best-effort), so failed and timed-out runs no longer report zero spend.

Testing

• Helper verified at runtime: unset -> 1800, 3600 -> 3600, 0 -> disabled, garbage -> warning + 1800. Both changed backend modules compile and import cleanly (no circular import).
• tsc passes; lint and formatting clean for the changed frontend file; store test suites pass.
• Reproduced the original failure with a long writing task aborted at exactly 1800s; with the token recording in place, an aborted run's history row carries the tokens consumed up to the failure.