SNRM2 precision

[dasergatskov]

There is a thread on OpenBlas
OpenMathLib/OpenBLAS#5503

which may be of some interest for you guys as well.

Currently, out of NETLIB, BLIS, OpenBlas, Intel MKL, Apple Veclib,
NETLIB and BLIS have low (but consistent with single precision),
and others have precision higher than eps(single) (assumingly by doing
summation in double precision).
I think it would be great if BLIS would do the same, but as a minimum
am curios about your opinion on this matter.

[devinamatthews]

I would suggest:

1. Add sdsnrm2 in line with What level of precision are users to expect of SNRM2 ? Reference-LAPACK/lapack#1180.
2. Add configure option to have snrm2 defined as sdsnrm2. Whether this is default on or not would need some more input.

[devinamatthews]

I guess I should say that if NETLIB is the only other outlier then we should probably not be doing that.