Expose _bbox_coordinates of BoundingBoxTree to users

[ordinary-slim]

Addressing #3599. Expose _bbox_coordinates of BoundingBoxTree the same way _x of Geometry is exposed. I added a test with point-tree collisions before and after a translational motion.

[ordinary-slim]

I've been profiling compute_collisions with points and I found some opportunities for speed-ups:

• Early break in point_in_bbox. Note that &= doesn't short-circuit and always evaluates the RHS so it's better to have a &= b && c in the same line rather than have it in two separate lines.
• Avoid calling BoundingBoxTree.get_bbox as it spends a good deal of time in copy_n. I wrote a small workaround as a lambda function in compute_collisions with points but I think it would be cleaner if get_bbox returned a span of const rather than a new copy.

In this snippet of code

from dolfinx import mesh, geometry
import numpy as np
from mpi4py import MPI
from line_profiler import LineProfiler

comm = MPI.COMM_WORLD
rank = comm.Get_rank()

def main():
    tdim = 3
    num_elems_side = 20
    domain = mesh.create_unit_cube(MPI.COMM_WORLD,
                                   num_elems_side, num_elems_side, num_elems_side,
                                   mesh.CellType.tetrahedron, dtype=np.float64)
    bbtree = geometry.bb_tree(domain, tdim, padding=0.0)
    rng = np.random.default_rng(0)
    num_points = 200000
    points = rng.random((num_points, 3))
    collisions = geometry.compute_collisions_points(bbtree, points)

if __name__=="__main__":
    lp = LineProfiler()
    lp_wrapper = lp(main)
    lp_wrapper()
    with open(f"profiling_rank{rank}.txt", 'w') as pf:
        lp.print_stats(stream=pf)

the line collisions = geometry.compute_collisions_points(bbtree, points) takes 2.4 seconds before these changes and 1.95 seconds after these changes.

[jhale]

Could you split these changes into two PRs? The algorithmic changes need a different type of review.

[garth-wells]

I don't think this is needed - use subspan.