Previous implementation used a lookup table of 200 bytes which, if not in L1 cache, could cause memory stalls and consumes 3-4 cache lines. This is suboptimal in real world scenarious. We are going to use a purely scalar version of merging 2/4/8 bytes together.

PiperOrigin-RevId: 533576619
Change-Id: Icd730d18536f7eb35979b62582f6edf86b786019