optimized fneg

[ZERICO2005]

I have optimized fneg, making the common case 3F faster. The main decision left is for inputs [+0.0f, +FLT_MIN * 2.0f), since we can make the case of input == 0.0f slightly faster (-1F + 1) at the cost of [+FLT_TRUE_MIN, +FLT_MIN * 2.0f) being slightly slower (+1F + 1)

Previous code:

    ; 13 bytes
    ; A != 0        :  8F + 3R + 0W + 2
    ; A:UBC == 0.0f : 11F + 6R + 3W + 2
    ; A:UBC != 0.0f : 14F + 6R + 3W + 2

Option 1 (optimize equally):

    ; 12 bytes
    ; A != 0        :  5F + 3R + 0W + 2
    ; A:UBC == 0.0f : 13F + 6R + 3W + 3
    ; A:UBC != 0.0f : 13F + 6R + 3W + 3

Option 2 (optimize input == 0.0f):

    ; 13 bytes
    ; A != 0        :  5F + 3R + 0W + 2
    ; A:UBC == 0.0f : 12F + 6R + 3W + 3
    ; A:UBC != 0.0f : 14F + 6R + 3W + 3