Draft: Performance optimization, bugfixes and cleanup

Good news up front: loading a benchmark file with 2M tets (no props) with bottom-up incidence generation goes from 18s down to 5s on my machine.

• Fix some dangling references
• Avoid lots of unnecessary copies
• reserve() vectors where sensible
• clean up code to make it more readable, e.g. new-style for-each loops, logic simplification

potentially breaking changes

• Some likely inconsequential API changes: add_{face,cell} now take std::vectors instead of references to them. This avoids copies (as the refgerenced vector was always copied) and should not need any changes in downstream code.

• adjacent_halfface_in_cell(hfh, heh) now assumes that the halfedge is part of the halfface's face. This is a reasonable assumption, but client code may have used it wrongly before and might break in that case. This was the case for a broken unittest.

TODO

• Look into CI failure
• make sure we reserve() when loading properties and measure results
• Bump version to 2.2, add changelog entry (see notes above)