fix(backend/executor): bound LLM calls and reap orphan node_execs (#13051)

## Why

A prod alert fired at 2026-05-08 02:07 UTC for graph execution
`51af771e-…` running for 24h+. Investigation surfaced two underlying
defects:

1. **No timeout on most LLM provider calls.** Only the Anthropic branch
of [`llm_call`](autogpt_platform/backend/backend/blocks/llm.py) sets
`timeout=600`. OpenAI, Groq, Ollama, OpenRouter, Llama API, AI/ML API,
and v0 are all unbounded. A stalled HTTP read parks the calling executor
thread indefinitely, which in turn parks the dispatch loop in
`_on_graph_execution`, which prevents the cleanup `finally` (and its
`task.cancel()` on running blocks) from ever running.

2. **Orphan `RUNNING` node_executions.** When a graph_exec is marked
terminal externally (e.g. via `stop_graph_execution`,
`activity_status_generator`, or operator action) but the in-flight node
tasks haven't propagated cancellation yet, those node_execs sit at
`RUNNING` forever — even after the graph row is `FAILED`.

In the prod incident, both defects compounded: the
`AITextSummarizerBlock` was hung in `oai_client.responses.create()` with
no timeout, the parent graph_exec eventually got marked `FAILED`
somewhere outside the executor, and the node_execs remained `RUNNING`
indefinitely while the executor pod sat in 24h
`terminationGracePeriodSeconds` waiting for them to drain.

## What

- New module-level constant `LLM_REQUEST_TIMEOUT_SECONDS = 600` in
`backend/blocks/llm.py`.
- `llm_call` wraps the provider dispatch in `asyncio.wait_for(...,
timeout=LLM_REQUEST_TIMEOUT_SECONDS)` and re-raises as `TimeoutError`
with the offending model in the message.
- Every provider SDK call now also receives
`timeout=LLM_REQUEST_TIMEOUT_SECONDS` for nicer error messages and
SDK-level fast-fail (Anthropic's existing literal `600` is now using the
constant for parity).
- `aiml_api` was using sync `openai.OpenAI` inside an async function,
blocking the event loop — switched to `AsyncOpenAI`.
- New `reap_orphan_node_executions()` in `backend/data/execution.py`:
- Finds RUNNING node_execs older than `min_age_seconds=300` whose parent
`AgentGraphExecution` is in `FAILED` / `COMPLETED` / `TERMINATED`.
- Marks them `FAILED` via `update_node_execution_status_batch` with
stats `{"error": "orphaned_after_graph_terminal"}`.
- Wired through `DatabaseManager` RPC, exposed in `backend.monitoring`,
and scheduled in the executor scheduler on the same cadence as
`report_late_executions`.

## How

- Belt-and-suspenders timeout strategy: `asyncio.wait_for` is the outer
guarantee; the per-SDK `timeout=` lets each provider raise its native
error shape if it can. `wait_for` cancels the inner task on timeout,
which propagates `CancelledError` through httpx and frees the
connection.
- The reaper is conservative: only acts on node_execs that are at least
5 min old AND whose parent is already terminal. That avoids racing the
brief window between graph terminal-update and node-cleanup during a
healthy shutdown. Limit caps each pass at 1000 rows.
- New tests:
- `backend/blocks/test/test_llm.py::TestLLMRequestTimeout` — verifies
the wrapper times out a hanging mock and that the OpenAI SDK receives
the `timeout=` kwarg.
- `backend/data/execution_reap_test.py` — verifies the reaper batches
the right node_exec ids, no-ops when no parents are terminal, and
constrains the prisma query.
- `backend/monitoring/late_execution_monitor_test.py` — verifies the
monitoring wrapper passes the configured `min_age_seconds` to the DBM
RPC.

## Out of scope

- Adding a `DiscordTrigger` webhook integration to replace
`ReadDiscordMessagesBlock`'s gateway-listener pattern (separate latent
bug; tracked as a follow-up).
- Making graceful shutdown actively cancel in-flight runs (intentional
24h grace per project convention; this PR just makes sure individual
blocks within a run can't hang indefinitely).

## Checklist

- [x] My code follows the [contribution
guidelines](https://github.com/Significant-Gravitas/AutoGPT/blob/master/CONTRIBUTING.md)
- [x] I have performed a self-review of my own code
- [x] I have commented my code where necessary
- [x] I have made corresponding changes to the documentation (n/a)
- [x] My changes generate no new warnings
- [x] I have added tests that prove my fix is effective