Slow ?trtrs implementation

[thrasibule]

openblas uses the reference implementation which is pretty slow since it uses ?trsm by default which is much slower than ?trsv in case the right hand side only has 1 column. This seems to come up regularly and surprise people (see here for a numpy discussion or here for the same discovery in julia). I noticed ?getrs has an optimized implementation which uses this simple trick (?trsv in case of column vector b, ?trsm otherwise), so I propose to do the same for ?trtrs. Does it make sense?