For block width >= 16, avx2 can further speedup the
TM_PREM intra prediction.

Function speedup on i7-6700:
Predictor  avx2 v. ssse3
16x8       ~1.6x
16x16      ~1.8x
16x32      ~1.9x
32x16      ~1.9x
32x32      ~1.9x

Change-Id: I62c20bd7628f52251b0c051b99a9b738ee44f7e6