[BugFix][CUDA] Support int4x2 and uint4x2 codegen (#3036)

* fix(cuda): support int4x2 and uint4x2 codegen (#2994, #2995)

* fix(cuda): handle packed int4 lane access

* fix(cuda): handle packed 4-bit x2 construction

* fix(cuda): load packed x2 gathers by logical index

* fix(cuda): validate packed int4 vector alignment

* fix(cuda): preserve packed vector load fallback