Solved New Zealand model ultra-slow on "Applying uncertainties"

[micheles]

In reality the problem was in .set_trt_smr because the cache on group_array(source_data, 'source') was not used :-( The speedup is huge, from 1 hour to 2 minute in the full calculation. In a reduce calculation I get

# before/after
| ncalls | cumtime | path                                |
|--------+---------+-------------------------------------|
|      1 | 825.102 | source_reader.py:399(_build_groups) |
|      1 |  53.062 | source_reader.py:399(_build_groups) |

Supersedes https://github.com/gem/oq-engine/pull/9687.

NB: the fake change on the CSV has nothing to do with the PR. Damned CSVs!