gh-8557 adds a hermitian=False kwarg to np.linalg.matrix_rank, which allows the use of a more efficient algorithm when the matrix is known to be hermitian.

Some other functions in np.linalg have similar considerations for this case, but require calling a different function for the hermitian case:

• eigvals / eigvalsh
• eig / eigh

(There's also fft vs hfft, but I don't want to touch that API if I can avoid it.)

The proposal: add hermitian keyword arguments to eigvals and eig, and encourage their use in place of eigvalsh and eigh.

Considerations: gh-8557 uses a simple boolean value, but @eric-wieser suggested allowing UPLO-style strings as well. This can be added to matrix_rank retroactively if desired, without breaking compatibility.