LATX, AVX: avoid loading overwritten VPMOVSX destinations

[luzeng87]

Summary

• allocate VPMOVSX destinations without loading their previous YMM contents
• preserve source upper-half state when source and destination differ
• add an x86-64 assembly regression covering aliased and distinct registers

VPMOVSX fully overwrites its destination. Loading the old destination can materialize upper-half state that the extension immediately replaces.

Validation

• clean LoongArch release build with GCC -O2 and LATX optimize O1
• native x86-64 assembly fixture: pass
• LATX JIT: pass
• LATX cold AOT: pass, non-empty AOT file generated
• LATX hot AOT: pass

[luzeng87]

Correctness review found that the production change is in translate_vaddsubpd(), while the PR and test describe VPMOVSX. Moving the allocation change to translate_vpmovsxxx() then fails the destination/source alias case in a clean LoongArch JIT, cold-AOT, and hot-AOT check. Closing this PR instead of merging a mis-targeted or unverified optimization.