Change probability initialization to use np.ones_like

[vivienr]

prior.sample() seems to always trigger the warning:

bilby WARNING : Prior sampling efficiency is very low, please verify its validity.

when using a prior with keys, without Constraints, and requesting more than 1000 samples. E.g.:

import bilby
prior=bilby.core.prior.PriorDict({'param1':bilby.core.prior.Uniform(minimum=5, maximum=50)})
samples=prior.sample(1001)

(And for some parallel workflows keeps triggering the warning).

This seems to stem from evaluate_constraints() which defaults to returning a single 1 when there aren't Constraints (instead of an array of ones), which is then used to count the number of valid samples (n_valid_samples), returning always 1, and thus triggering the warning once one asks for more than 1000 samples (if n_tested >= 1e3 and efficiency < 1e-3: in check_efficiency()).

The solution proposed here is to replace evaluate_constraints()'s default prob to be an array of ones like the first values in out_samples.

[ColmTalbot]

Looks good to me