Specialization of Normed{UInt16} -> Float32 conversions #373

While working on the PR #372 benchmarks, I realized that some Normed{Uint16}->Float32 conversions hadn't been optimized (or specialized). (cf. Issue #129, PR #138)

Compared to seven years ago, Julia has become more flexible in terms of broadcasting and inlining.
Furthermore, even taking into account the previous constraints in Julia, the following should still be effective.

    f == 10 && return Float32(Int32(x.i) * 0xcd) * @f32(0x500005p-40)

Also, I think the f==8 case, which is used relatively often, is worth optimizing.

    if f == 8
        s = x.i <= 0xff00
        f32m = Float32(x.i - (s ? 0x0000 : 0xff00))
        return f32m * @f32(0x101p-16) + f32m * @f32(0x1p-24) + (s ? 0f0 : 256f0)
    end

Although the handling of fma() has improved compared to seven years ago, the performance penalty in the absence of hardware support is still significant, so we should avoid using it to improve speed rather than accuracy.

It might be okay to use Float64 operations, but I don't have a thorough understanding of GPU compatibility or performance with Float64.