print floats more accurately

[regehr]

earlier today I was working with this example:

define i1 @src(i16 %x, i16 %y) {
    %x1 = uitofp i16 %x to float
    %y1 = uitofp i16 %y to float
    %sub = fsub float %x1, %y1
    %abs = tail call float @llvm.fabs.f32(float %sub)
    %ext = fpext float %abs to double
    %cmp = fcmp olt double %ext, 1.000000e-15
    ret i1 %cmp
}

define i1 @tgt(i16 %x, i16 %y) {
    %cmp = icmp eq i16 %x, %y
    ret i1 %cmp
}

Alive verifies it but prints the float as 0.000000. with this patch we print the float as 1.0000000000000001E-15

also works for other float types, let me know if you want tests on this

[nunoplopes]

does this need to check for NaN?

[nunoplopes]

I'm very surprised we would need this patch. Either bitcastToAPInt is not working correctly, or something is off with the printing in constant.cpp. Representing the bit pattern of the number should be the most precise way of storing it.