One of the terms in m[0][0] and m[1][1] was incorrectly negated from an
incomplete factorisation when I transcribed the formula.
Also reintroduces the axis-rotation tests which inadvertantly had zero
magnitudes.
comparison of floating point values is inexact and there isn't an easy
way for providing an ULP delta via the TAP interface. morever, the
general expected error is too large to reliably express in ULPs.
