[n, k//2] int4pack to [n, k//8] w/o transpose

[ZhiweiYan-96]

Motivation

1. Original should be [n, k//2] int4x2, target should be [n, k//8] int4x8, no need to transpose, so the indexing logic is changed
2. [Need to be decided] We should not consider big endian/small endian here right? I think int4 should be packed in easiest way, aka, just right shift per bits.

[airMeng]

although might never happened, but could you handle K_ not divided by 8?

[ZhiweiYan-96]

good point, let me add an assert