Optimize gravity forward of prism layer

[santisoler]

Re-implement the gravity method of the prism layer: rewrite the forward modeling computation of the gravity fields of a prism layer through a custom Numba function that iterates directly over the 1D horizontal coordinates of the layer. Replace the old implementation that used to build the whole set of prism boundaries and call the prism_gravity function. This new implementation doesn't require to store all those prisms in memory, significantly lowering the memory requirements, especially for large problems. Remove obsolete private function and method in layer.py. Also, reorganize prism forward modelling into a submodule: group together prism-related pieces of code inside a new harmonica._forward/prisms module, organized in a similar way as the one for ellipsoids.

Relevant issues/PRs:

This issue has been found by @aguspesce, who noticed a significant amount of memory being used while forward modeling a large prism layer.

[santisoler]

I benchmarked this branch against main (e60e717) with this script, running the forward of a big prism layer on single computation point (just to speed up things, since the memory issue is coming not from the number of observation points, but from the allocation of prisms boundaries):

import numpy as np
import harmonica as hm
import bordado as bd

# Build prism layer
region = (-10e3, 10e3, -10e3, 10e3)
shape = (5000, 5000)
non_dimensional_coords = 100.0
easting, northing, surface = bd.grid_coordinates(
    region=region, shape=shape, non_dimensional_coords=non_dimensional_coords
)

properties = {"density": np.ones_like(easting)}
prism_layer = hm.prism_layer(
    (easting, northing), surface=surface, reference=0, properties=properties
)
print(prism_layer)

# Build observation point
coordinates = (0.0, 0.0, 500.0)

# Benchmark performance
prism_layer.prism_layer.gravity(coordinates, field="g_z")

I ran this script using:

command time -v python script.py

On main I obtained a peak usage of memory of:

Maximum resident set size (kbytes): 4505612

While on this branch (eeb94dd) I get:

Maximum resident set size (kbytes): 1877456

This shows a significant reduction in memory usage.