Fix memory occupation issues in scenario_risk

[Michele Simionato]

This is a companion of https://bugs.launchpad.net/openquake/+bug/1132017

Having identified the issue, one should be able to fix it, possibly by changing the query that was reclaiming all that memory, or by simplifying the table structure, or both.

[Michele Simionato (michele-simionato)]

The fix requires changing both the gmf_scenario table and the distribution logic. The gmf_scenario table must have few rows, possibly with a long array for each row, to keep a good performance even for large site collections. Currently instead the gmvs are split in chunks (~30 chunks per location). The proposal is to distribute not according to the realizations, but in blocks of sites. For instance each task could compute all the realizations of a given GMF for a block of sites (say 1000 sites per task).