multivariate_normal warns about non-PSD covariance for float32 inputs

[mbillingr]

Passing a clearly positive definite covariance matrix with float32 data type causes the warning

RuntimeWarning: covariance is not positive-semidefinite.

Using the same matrix with float64 does not raise the warning.

Steps to reproduce:

import numpy as np
cov = np.array([[1, 0.1], [0.1, 1]], dtype=np.float32)
np.random.multivariate_normal([0, 0], cov)

This was raised as a question at StackOverflow.

I suspect that the behavior may be related to the way the function's tol argument is passed to both rtol and atol in psd = np.allclose(np.dot(v.T * s, v), cov, rtol=tol, atol=tol). The default of tol=1e-8 is much lower than np.isclose's default for rtol=1e-5.