x86: Fix overflows in inverse identity transforms

Fixes #303 (closed).

Functions with the largest impact (although identity transforms are a negligible fraction of overall runtime to begin with):

identity_identity before after
16x16_ssse3:      165.7  185.0
16x16_avx2:        94.6  102.4
32x16_ssse3:      304.0  341.1
32x16_avx2:       161.1  176.7