Reduced memory consumption in the realization weights

[micheles]

For instance the EUR model has 302_990_625 realizations and trying to build the realizations causes cole to run out of memory (512 GB) after 2.5 hours. Now building the weights takes only 25 minutes and requires < 100 GB of RAM. However, the performance is dominated by full_lt.get_rlzs_by_gsim which is pretty slow and memory consuming. Here is the time spent in full_lt.init() for New Zealand with ~1M realizations:

| old vs new       | time_sec | memory_mb | counts |
|------------------+----------+-----------+--------|
| building full_lt |     54.5 |     639.6 |      1 |
| building full_lt |     52.9 |     365.8 |      1 |

The memory is nearly halved.