[Matching] Fix L1 metric

- Restrict keyword is useless if using explicit SIMD instruction
(and cause) trouble with data alignment
- implicit vectorisation is on par with explicit vectorisation
but we keep both for now
- Timings - explicit SSE2 L1 kernel Vs L2 kernel
Linux Computer / 24 threads / Approx. 2 Ghz.
. Cabane JB - 895 pairs / 50k feats per images on average:
	HNSWL1 209 s.
	HNSWL2 243 s.
. Chapelle JB - 10K pairs / 50 feats per images on avrage:
	HNSWL1 2018 s.
	HNSWL2 2300 s.
	CASCADE 2526 s.
- Timings - explicit AVX2 Kernel for both L1 and L2
Windows Computer / 8 threads / Approx 2,5 Ghz.
. Gom. - 130 images (exhaustive matching)
	CASCADE 229 s.
	HNSWL2 389 s.
	HNSWL1 411 s.
	ANNL2 (no SIMD or implicit) 900 s.
Note: here CASCADE is faster but produce more FP than HNSW and FLANN
- Fix the unit test for L1 kernel