`rhat_nested` returns `nan` when the number of draws is one (or less than four)

[pipme]

import numpy as np
from arviz_stats.base import array_stats

rng = np.random.default_rng()
samples = rng.normal(size=(4, 1, 2))
axis = {"chain_axis": 0, "draw_axis": 1}
print(array_stats.rhat_nested(samples, (0, 0, 1, 1), **axis))

This outputs:

INFO:arviz_stats.base.stats_utils:Shape validation failed: input_shape: (4, 1), minimum_shape: (chains=2, draws=4)
INFO:arviz_stats.base.stats_utils:Shape validation failed: input_shape: (4, 1), minimum_shape: (chains=2, draws=4)

[nan nan]

According to the original paper on nested rhat (https://arxiv.org/abs/2110.13017), there should not be such a restriction on the minimum number of draws. Any reason for this restriction in arviz_stats?

Thanks!

[aloctavodia]

This is mostly inherited from the "standard" Rhat. For instance, for the "rank" method (default), we split the chain in two and then compute the variance (with 1 dof), so we require a minimum of 4 draws. For the "identity" method (and maybe just withing the nested rhat), we may want to relax this... as this will closely match the paper and possible posterior code... @zaxtax What do you think?