Further investigation into Windows ARGB

[sudara]

I wasn't able to get better performance than baseline (gin) for ARGB version on Windows:

Routes tried:

• Various IPP methods like separated convolution. These were faster, but broke down at higher radii, probably because under the hood they are a 2D kernel.
• The "rotated" vector implementation (where the queue is vertical) with FloatVectorOperations — this was consistently slow
• Rotated vector implementation in IPP — again, performance seems to be worse than 4x the single channel, no matter how it's implemented (looping around the channels, allocating 4 channels worth of queue/temp storage, etc)

Things to try:

• Investigate whether Alpha needs to be blurred at all, given the pixels are premultiplied.
• Custom dot-product with the kernel
• The "journey of a pixel" sliding kernel idea from my blog