Convert CellDetector to numba

This is the final peice in the puzzle for https://github.com/brainglobe/cellfinder-core/issues/89, converting the CellDetector class to numba. Some comments for context:

Erorrs raised in numba code must have a message determined at compile time, which means I had to re-write several to avoid format strings.
connect_eight() isn't touched by this PR. It isn't tested by our test suite, and I think it's not possible for a user to run that method anyway. If we want to keep connect_eight(), I'd advocate for doing a numba translation and adding tests at the same time in a new PR (I've opened https://github.com/brainglobe/cellfinder-core/issues/111 for this)
I had to merge StructureManager into CellDetector (it's not possible to use one jitclass in another jitclass)

After this PR, running benchmarks_3d.py gives:

4.304 <module>  filter_3d.py:1
└─ 4.304 VolumeFilter._run_filter  cellfinder_core/detect/filters/volume/volume_filter.py:104
   ├─ 4.091 wrapper  numba/experimental/jitclass/boxing.py:59
   │     [17291 frames hidden]  numba, .., llvmlite, colorama, _colle...
   └─ 0.213 BallFilter.walk  cellfinder_core/detect/filters/volume/ball_filter.py:129
      ├─ 0.164 CPUDispatcher._compile_for_args  numba/core/dispatcher.py:388
      │     [923 frames hidden]  numba, llvmlite, colorama, _collectio...
      └─ 0.049 _walk  cellfinder_core/detect/filters/volume/ball_filter.py:197

Confirming that connect_four runs quicker than 0.049 seconds with the conversion to numba. This is similar or better than the original Cython implementation, where it also didn't show up on the pyinstrument trace: https://github.com/brainglobe/cellfinder-core/issues/89#issue-1604899152

Fixes https://github.com/brainglobe/cellfinder-core/issues/89.

brainglobe / cellfinder-core

Convert CellDetector to numba #112