LATX, AVX: streamline integer narrowing and rounded multiply

[luzeng87]

Summary

• use LASX saturating narrowing directly for VPACKUSWB and VPACKUSDW instead of pre-clamping negative lanes
• avoid unnecessary temporaries and copies for signed and unsigned packing while preserving register-overlap cases
• replace the expanded VPMULHRSW rounding sequence with the matching LSX/LASX rounded arithmetic shift

Validation

• clean 3A6000 build: passed
• signed-to-unsigned saturation boundary cases: passed
• VPMULHRSW edge cases and 100,000 deterministic input sets: passed
• JIT, cold AOT, and hot AOT: passed with non-empty AOT files
• git diff --check: passed

[luzeng87]

Superseded by grouped AVX lowering PR #448. The original commit remains separate inside that PR for review.