Rewrite the classical calculator as a map-reduce job

[Michele Simionato]

Recently we proved that passing large amounts of data via celery/rabbitmq is as efficient as storing them in the database, if not better. Then the current approach used in the classical calculator, in which partial results are stored in the database, locking the hazard_curve_progress table and thus serializing the computation, has no justification. Hundreds of lines of code can be removed and the calculator can be made both simpler and more efficient by using a standard map-reduce approach in which the arrays are kept in memory and aggregated in the controller node, not in the database.

Bonus point: one can also remove the special case complication of point sources.

[Michele Simionato (michele-simionato)]

https://github.com/gem/oq-engine/pull/1337