Add float32 compatibility to KMedoids

[TimotheeMathieu]

I implemented the changes suggested by @rth in PR#83 for dtypes in kmedoids.
In Clara, the dtype changes can make a consequent speedup but for kmedoids, I don't know why but there is no speedup:

Here is a small benchmark :

• KMedoids

	wall_time	cpu_time	peak_memory
('build', 'float32')	0.178639	0.49873	1.14062
('build', 'float64')	0.166929	0.396058	0.0859375
('heuristic', 'float32')	0.0497368	0.255722	0.0117188
('heuristic', 'float64')	0.0783268	0.250505	60.8164
('k-medoids++', 'float32')	0.0602323	0.261578	0.00390625
('k-medoids++', 'float64')	0.0871284	0.284815	61.0117

• CLARA

	wall_time	cpu_time	peak_memory
('build', 'float32')	0.944996	3.84636	0.160156
('build', 'float64')	0.794827	5.21157	1.21875
('heuristic', 'float32')	0.956299	4.54045	0
('heuristic', 'float64')	0.803466	5.22568	0
('k-medoids++', 'float32')	0.940615	6.23457	0
('k-medoids++', 'float64')	1.14663	6.11493	0.00390625

For Clara we can design settings in which the difference between 64 and 32 bit is quite large. Here I used n_sample = 200 000 in dimension 100 for CLARA and sampling_size = 200 . I used n_sample = 2000 for KMedoids.

Code

import numpy as np
import neurtu
from sklearn_extra.cluster import KMedoids, CLARA

X = np.random.normal(size=(100_000, 100))
X_32 = X.astype(np.float32)

def make_experiment(dtype, init):
    if dtype == 'float32':
        X2 = X_32
    else:
        X2 = X
    #km = CLARA(init=init, sampling_size= 50, n_clusters=9)
    km = KMedoids(n_clusters = 9, init = init)
    km.fit(X2)


def cases():
    for init in ['build', "heuristic", "k-medoids++"]:
        for dtype in ['float32', 'float64']:
            tags = {'init' : init, "dtype": dtype}
            yield neurtu.delayed(make_experiment, tags=tags)(dtype, init)
        
        
bench = neurtu.Benchmark(wall_time=True, cpu_time=True, peak_memory=True)
df = bench(cases()) # 2

[rth]

Thanks could you parametrize an existing test to run both on float64 and float32 input? and maybe check that the output of transform is of the same dtype. Otherwise LGTM.

Something like,

@pytest.mark.parametrize('dtype', [np.float64, np.float32])
def test_...(dtype):
    X_input = X_input.astype(dtype)
    ..