Saving memory in conditioned scenarios

[micheles]

By using shared memory, which means conditioned scenarios cannot run on clusters, which is fine, since they would be terribly inefficient without shared memory due to the data transfer. The improvement is HUGE.
The following calculation (taken from the Gorkha scenario, job.zip) which was not running at all with 48 GB of RAM, with errors like

[2024-05-16 07:38:33 #12498 INFO] event_based  40% [20 submitted, 0 queued]
Process SpawnPoolWorker-57:
Traceback (most recent call last):
  File "/usr/lib64/python3.11/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/usr/lib64/python3.11/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/usr/lib64/python3.11/multiprocessing/pool.py", line 114, in worker
    task = get()
           ^^^^^
  File "/usr/lib64/python3.11/multiprocessing/queues.py", line 367, in get
    return _ForkingPickler.loads(res)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
_pickle.UnpicklingError: invalid load key, '\xc3'.

Now runs in 2 minutes on the aristotle-beta machine with minimal data transfer:

| calc_2, maxmem=30.9 GB    | time_sec | memory_mb | counts |
|---------------------------+----------+-----------+--------|
| EventBasedCalculator.run  | 136.2    | 40.0      | 1      |
| total event_based         | 99.6     | 31.8      | 34     |
| updating gmfs             | 49.2     | 0.0       | 400    |
| instantiating GmfComputer | 44.7     | 0.0       | 100    |

$ oq show job_info
| task        | sent                                                | received | mean_recv |
|-------------+-----------------------------------------------------+----------+-----------|
| event_based | stations=19.99 MB cmaker=341.43 KB proxies=40.87 KB | 11.28 MB | 339.76 KB |

There is no other way if we want to save disk space and avoid the pickle limitations. Can be considered part of #9387.

[raoanirudh]

Great!