… resurrection

When a crawler run was resumed from a persisted state, two issues inflated
the reported `crawler_runtime`:

- `Statistics.__aenter__` started the periodic logger before updating
  `crawler_last_started_at`, so the first statistics log of the resumed run
  computed the runtime from the previous run's start time and included all
  the downtime between the runs (later stats were computed correctly, which
  made the value appear to "jump back").
- When the state came from a run that did not finish cleanly (migration,
  abort), `StatisticsState.model_post_init` computed the accumulated runtime
  offset using the current time, permanently baking the downtime between the
  runs into all subsequent statistics. The offset is now computed from the
  moment the state was last persisted instead.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VMHwDx87jV4rnu5LkbpNxy

The first periodic statistics log must report exactly the runtime
accumulated by previous runs: zero for a fresh start, the previous total
for a resumed run. `Statistics.__aenter__` now resets the run timestamps
before starting the periodic logger and sets the new start time after the
initial log entry is printed.

The accumulated runtime is now also restored from the persisted
`crawlerRuntimeMillis` value instead of being reconstructed from the last
run's timestamps, so runs before the most recent one are no longer lost
from the total, and a state persisted mid-run keeps the correct runtime.
The timestamp-based reconstruction is kept as a fallback for states
persisted by older versions. `crawlerRuntimeMillis` is also serialized as
milliseconds now, matching its name and the other `*Millis` fields,
instead of an ISO 8601 duration string.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VMHwDx87jV4rnu5LkbpNxy

… on Python 3.10

`datetime.fromisoformat` does not accept the 'Z' suffix until Python 3.11,
which failed the resumed-run test on the 3.10 CI jobs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VMHwDx87jV4rnu5LkbpNxy

PR #2204 restores the accumulated crawler runtime from the persisted
'crawlerRuntimeMillis' value whenever the key is present, and falls back
to timestamp-based reconstruction only when the key is absent. However,
every crawlee release before v1.1.0 persisted 'crawlerRuntimeMillis' as
a plain model field that nothing ever updated, so those states always
contain the key with value 0 (verified against crawlee 1.0.2 and 0.6.0
from PyPI). Resuming from such a state therefore restores a zero runtime
offset and, because __aenter__ now also resets the run timestamps, the
entire previous runtime is silently dropped instead of being
reconstructed from the timestamps as before.

The clean-shutdown test passes on master and fails on the PR head; the
unclean-shutdown test documents that the stats_persisted_at fallback is
unreachable for real persisted states of any released version.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BPQjAuUxxHHFv7EqAiSZrM