revert to multiplication for radian-degree conversion

[simonbyrne]

I tried to be a bit too clever (or not clever enough) in 4a5161d by splitting the multiplication when converting between radians and degrees: although this method is correct for 30 degrees, on average it actually gives worse rounding:

julia> X = rand(1000);

julia> sum(abs(Float64[degrees2radians(i) - float64((big(pi)/180)*i) for i = X]))
3.2187421402252864e-16

julia> sum(abs(Float64[(pi/180)*i - float64((big(pi)/180)*i) for i = X]))
1.6530017372257122e-16

This commit reverts to using ordinary multiplication (and so sind(30) == 0.49999999999999994).

It is possible to do multiplication by irrational numbers with correct rounding, based on splitting the floats into half-precision (see Dekker (1971) for details of half-precision methods), but these methods require more computations. For example:

dpb = big(pi)/180
dp  = float64(dpb)
dph = signif(dp,26,2)
dpt = float64(dpb-dph)

# based on Veltkamp (1968), see Dekker (1971), p. 234
function degrees2radians(x)
    xh = signif(x,26,2)
    xt = x - xh
    z = x*dp
    z + ((((xh*dph - z) + xh*dpt) + xt*dph) + xt*dpt)
end

This approach could be used for any non-float-representable number, however it requires 11 floating point operations (plus a truncation), so is probably not worth it just to get the final digit correct.

[simonbyrne]

For future reference: this can be done in 6 flops + a truncation (which internally has a couple of flops, but could get optimised since they're powers of 2).

dpb = big(pi)/180
dp  = float64(dpb)
dph = signif(dp,26,2)
dpt = float64(dpb-dph)

# simplified
function degrees2radians(x)
    xh = signif(x,26,2)
    xt = x - xh
    xh*dph + (xt*dph + x*dpt)
end

However we still have to deal with the fact that we have to round before calling sin (see discussion of #4112).