Is ``apollon.tools.fsum`` really necessary?
A small test using the below snippet showed that the error has magnitude of about 1e-12.
out = []
for i in range(100):
samples = np.random.randn(1000, 1000)
for j in range(10):
x, y = np.random.randint(0, 1000, 2)
samples[x, y] = np.random.randint(0, 1000)
out.append(samples.sum() - fsum(samples)))