incorrect `diagnostic.l_ratio_bounds` for `has_variance=True`

[jorenham]

The strict bounds should be divided by the L-scale, not by the upper bound of the L-scale. This results is bounds that are too tight.

E.g. the strict untrimmed L-moment bounds for $r \ge 3$ are $\sqrt{2r - 1} |\lambda_r| \le \sigma$, so the corresponding L-ratio bouds are $\sqrt{2r - 1} |\tau_r| \le \sigma / \lambda_2$.