Use SSE instructions for point rotation in ORB extraction

[shinsumicco]

Fully use SSE instructions for point-rotation before comparing two points in ORB extraction.

#define COMPARE_ORB_POINTS(shift) \
        (_point_pairs = _mm_load_ps(orb_point_pairs + shift), \
         _mul1 = _mm_mul_ps(_point_pairs, _trig1), \
         _mul2 = _mm_mul_ps(_point_pairs, _trig2), \
         _vs = _mm_hadd_ps(_mul1, _mul2), \
         _vi = _mm_cvtps_epi32(_vs), \
         _mm_store_si128(reinterpret_cast<__m128i*>(ii), _vi), \
         center[ii[0] * step + ii[2]] < center[ii[1] * step + ii[3]])

[shinsumicco]

Tracking times are reduced by 2-3%.

(measured using KITTI 05 monocular seq., without OpenMP)
w/o SSE: 0.034165[s/frame]
w/ SSE: 0.0331221[s/frame]