Improve performance

[bl-young]

This resolves some challenges I was having generating the 2023 baseline from the current develop branch. It includes two main changes that have negligible effect on the final datasets:

1. In combinator.py, adds _build_flowmapping_small_for_merge() for generating a more manageable flow mapping file which preserves intentional eLCI 1:many splits (e.g. sand/gravel) but doesn't blow up the mapping when pulling in the rest of the fedefl (via Unexpected mis-matched flows in eLCI.csv #274). This helped me avoid Memory Errors

2. In generation.py, replace groupby(...).apply(_combine_sources) with vectorized helpers to improve performance.