[core] Faster ground profile querying.

[duburcqa]

• [core] Replace MurmurHash3 by xxHash32 which is faster.
• [core] Make gradient computation optional for heightmap functions.

Fixes #666

The snippet mentioned here #799 (comment) is now taking 6us instead of 13us on my machine (3us if all points are co-located, 11us is they are extremely far away from each other). I don't think it is possible to do any better at this point without significant refactoring, but I think it is good enough now.

In particular, this meaningful snippet runs in 17us on my machine:

import numpy as np
from jiminy_py.core import random_perlin_ground, query_heightmap

x_dim, y_dim = 20, 20
x_grid, y_grid = (np.linspace(-1.0, 1.0, i) for i in (x_dim, y_dim))
vertices = np.stack((
    np.tile(x_grid, y_dim),
    np.repeat(y_grid, x_dim),
    np.empty((x_dim * y_dim,))
), axis=0)

ground_fun = random_perlin_ground(wavelength=1.0, num_octaves=8, seed=42)
%timeit query_heightmap(ground_fun, vertices[:2], vertices[2])