ENH: vectorize `cov`

[mdhaber]

cov currently expects an array with at most two dimensions. Given a single $m$-dimensional sample composed of $n$ observations as an array with shape (m, n), it produces one $m \times m$ covariance matrix as an array of shape (m, m).

It would be useful if cov were vectorized operate on arrays of arbitrary dimensionality as a batch. Given a batch of samples as an array with shape batch_shape + (m, n), it would produce a batch of covariance matrices as an array of shape batch_shape + (m, m).

This would be consistent with other standard functions like matmul. matrix_transpose, and many linalg extension functions.

(I'd have an immediate use for this vectorizing scipy.stats.epps_singleton_2samp.)

[lucascolley]

SGTM!