Fix bug checksum failures

This commit fixes the bug where multiple realizations try to read checksum
at the same time, leading it to grind to a halt.
The issue was fixed by moving the forward_model_ok_lock up a level, so
that only one realization can call `verify_checksum` at a time. This
also solved the issue where jobs would be shown as stuck in pending even though
all forward model steps had completed. This was probably due to
checksums all being checked and verified at the same time, which can be
IO intensive.

src/ert/scheduler/job.py

@@ -150,9 +150,9 @@ async def run(
                 break
 
             if self.returncode.result() == 0:
-                if self._scheduler._manifest_queue is not None:
-                    await self._verify_checksum()
                 async with forward_model_ok_lock:
+                    if self._scheduler._manifest_queue is not None:
+                        await self._verify_checksum()
                     await self._handle_finished_forward_model()
                 break