Integrate fastcore functions

[schlegelp]

This is a work-in-progress PR to integrate functions re-implemented in Rust via navis-fastcore.

List of functions that will now use fastcore under-the-hood if available:

• geodesic_matrix(): 20x faster (note that this will also speed up a number of downstream functions such as longest_neurite)
• prune_twigs(): 4x faster
• strahler_index(): 20x faster
• synapse_flow_centrality(): 4.5x faster (but importantly much more memory efficient)
• segments_to_coords(): 3.5x faster (this function is a major bottleneck during plotting)
• TreeNeuron.cable_length property: 2x faster
• _connected_components(): 1.4x faster

Potential TODOs:

• [ ] skeletonization (bigger project)
• [ ] NBLAST (protoype working but oddly still slower than implementation in navis)
• [x] pin fastcore version

Misc:

• this PR also fixes a couple bugs encountered in the progress:
  • fixed synblast multiprocessing (pickling) issue
  • fixed issue with ax.add_collection3d in plot2d (introduced with in recent matplotlib version)
  • single disconnected nodes will now be ignored during plotting when radius=True (caused error previously)
  • neuPrint interface: fetch_mesh_neuron() will now ignore lod parameter if not supported by mesh source (e.g. for legacy meshes); previously this caused an error
• fixed a bunch of deprecation warnings e.g. from numpy, igraph and matplotlib
• subset_neuron now sets ._soma to None if the soma is lost during subsetting

[schlegelp]

There is one remaining failing test: strangely, MovingLeastSquaresTransform.xform return different values depending on whether I run it locally on my machine or whether it runs on Github. Since this problem is unrelated to this PR, I will ignore it and go ahead with the merge.