ENH: Cast covariance to double in random mvnormal

Fix Wald Docstring to reflect actual restriction on parameters
Add a test to ensure these are enforced for scalar inputs

Cast the covariance in the multivariate normal to double
so that the interpretation of tol is cleaner.

closes #10839

Add a check for NaN probabilities in random.choice

closes #11250

Before this fix, np.random.weibull(a=0) often returned inf (and
in theory could have returned 1).  It should only return 0.

Closes gh-12371.

BUG: Make `random.shuffle` work on 1-D instances of `ndarray` subclasses

Alter test to check if arrays are the same to avoid in-place of some
array-like objects

closes #11975