Is your feature request related to a problem? Please describe.
Even if the original input matrices were given as arrays, each of them is first turned into a LinearOperator:

    A = _makeOperator(A, (n, n))
    B = _makeOperator(B, (n, n))

When lobpcg is requested to compute > 20% of eigenvalues, it simply switches to eigh, but before that needs to recover the matrices back:

        A_dense = A(np.eye(n, dtype=A.dtype))
        B_dense = None if B is None else B(np.eye(n, dtype=B.dtype))

        vals, vecs = eigh(A_dense, B_dense, eigvals=eigvals,
                          check_finite=False)

which is a dramatic waste.

Describe the solution you'd like
Unless original input matrices were given as LinearOperator's, avoid

        A_dense = A(np.eye(n, dtype=A.dtype))
        B_dense = None if B is None else B(np.eye(n, dtype=B.dtype))

when computing > 20% of eigenvalues and call eigh directly on the original matrices.