Add caching of ranks for computing RSA with Spearman

[gydis]

I have added a cache_ranks flag to rsa_gen and rsa_array to enable caching of ranks for model RDMs when computing RSA with Spearman. For now, the ignore_nan=True case is ignored and an error is raised when cache_ranks is used with it.

[wmvanvliet]

This is a really neat little speedup!

[wmvanvliet]

would this also word for the tau-a metric?

[wmvanvliet]

To give an indication of the speedup of this PR, here is a speed test:

import mne_rsa
import numpy as np
from tqdm import trange

rng = np.random.default_rng(1)

# 10 model RDMs
model_rdms = [rng.random(1_000) for _ in range(10)]


# 100_000 data RDMs
def gen_data_rdms(n=100_000):
    """Generate the same RDM over and over again."""
    data_rdm = rng.random(1_000)
    for _ in trange(n):
        yield data_rdm


# RSA the 10 model against the 100_000 data RDMs
for _ in mne_rsa.rsa_gen(gen_data_rdms(), model_rdms):
    pass

On my analysis machine, using a single core, before this PR, this took about 13 minutes to complete. After this PR it takes a single minute.

[wmvanvliet]

thanks @gydis!