LATX, AVX: avoid redundant vector copies and lane inserts

[luzeng87]

Summary

• write VEXTRACTF128 register results directly to the destination
• handle VINSERTI128 source/destination overlap without an unconditional temporary and final copy
• remove redundant upper-lane inserts after full-width VPBROADCASTB/W/D operations

Validation

• clean 3A6000 build: passed
• explicit VEXTRACTI128/VINSERTI128 overlap cases: passed
• 256-bit byte, word, and doubleword broadcasts: passed
• JIT, cold AOT, and hot AOT: passed with a non-empty AOT file and empty stderr
• git diff --check: passed

[luzeng87]

Superseded by grouped AVX lowering PR #448.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