fix(api): widen the scheduled-run window in the in-flight cancel test (#2649)

Fixes #2648.

`update state while run in flight` schedules a run 2 seconds out and
assumes the create, update-state, and cancel round trips finish inside
that window. Since the GitHub runners moved to Node 26.5.1 (undici
8.9.0), client-side fetch dispatch under the suite's concurrent load can
stall ~1.5s, the cancel misses the window, and the run ends `"error"`
instead of `"interrupted"` — every CI run since Aug 1 is red on 26.x.
Full analysis with the undici A/B and `diagnostics_channel` trace is in
the issue.

Scheduling the run 5 seconds out fits the sequence even with slow
dispatch. The happy path cancels the run before it starts, so the test's
duration is unchanged.

Verified locally on unmodified logic: 3/3 failing runs on Node 26.5.1
before the change, 3/3 green after, still green on 26.5.0.

Not covered: the underlying undici 8.8.0 dispatch regression (needs a
standalone repro before an upstream report) and the question of whether
cancelling an already-running run should end it as `"interrupted"` —
both noted in the issue.