Optimizing multiplication for Normed

[kimikage]

The current implementation of multiplication for Normed uses floating point arithmetic.

The following three points are important in adding checked_mul:

1. Unlike checked_add/checked_sub, FixedPoint multiplication and rawtype multiplication are different.
2. N0f8, N0f16 and so on, do not cause overflow in multiplication.
3. Float-->Normed conversion code may throw an exception.

Therefore, the optimization for at least N0f8 and N0f16 is important.

[kimikage]

See also: #125 (comment)

However, reducing the bit width (i.e. use of UInt16) may be advantageous for SIMD loops.

function Base.:*(x::N0f8, y::N0f8)
    m = UInt16(x.i) * UInt16(y.i)
    z = m + ((m + 0x80) >> 0x8) + 0x80
    reinterpret(N0f8, unsafe_trunc(UInt8, z >> 0x8))
end