Should `trysl!` throw a `LAPACKError` for positive values?

[kshyatt]

According to the NETLIB docs, trsyl returns a value of 1 for the error code when A and B in the Sylvester equation have close eigenvalues and so perturbed values had to be used. Is that really something that should throw a full error rather than a warning?

[araujoms]

EDIT: I did some numerical experiments, and could only trigger the warning when the solution was in fact incorrect, so I think a full error is appropriate.