Gaussian blur

[awxkee]

This is some initial implementation of gaussian blur, at the moment, it is more too have a look rather than merge.
This implementation is ~x20+ faster than current, faster than fast blur until radius goes reasonable huge (~50+) and will be probably faster than any other blur implementation (stack blur etc ) on low radius, with rayon it should even more effective, parallelization level of this algorithm is very high.

Here is some concerns:

• Old implementation considers sigma something different, more like radius base instead of flattening level
• I'm not sure how to handle GenericImageView without explicit going unsafe; and even with unsafe I'm not sure I know best way to get direct typed buffer without copying.
• Loops are hardly unrolled what is not common practice in this crate, and this is important.
• This unrolling not really good adopted for x86 with low registers bank.
• Loops unrolled in current implementation with target u8 storage type in mind, for other types performance variance is higher
• It is named filter_1d, but at the moment only kernels with symmetry is handled properly. Are there some point to finalize it?
• This algorithm requires more memory than previous
• Some beautiful types as ColorGroup also is not common practice in this crate