LATX: optimize AVX scalar physics patterns

[luzeng87]

Summary / 变更说明

This PR depends on #437. Until #437 is merged, GitHub also shows its prerequisite commit in this PR; the GB603-specific changes are the final two commits.

• Write common VEX.128 results directly to the destination register and reduce redundant shuffle/copy operations.
• Fold adjacent VCOMIS*/VUCOMIS* plus JCC sequences while preserving ordered and unordered floating-point behavior.
• Fold repeated scalar additions and the six AVX three-lane sum groups found in Geekbench 7 workload 603.
• Match instruction and register relationships only; there is no program-address check and no Geekbench-specific runtime switch.
• Reject the sum3 pattern unless all skipped result registers are overwritten before any later use.
• Add native x86, LAT JIT, cold-AOT, and hot-AOT semantic tests for every new pattern.

This change does not modify the AOT cache identity or enable LAT --optimize-O2.

Performance / 性能

3A6000-25G, Geekbench 7 geekbench_avx2, workload 603, GCC 8.3.0 -O2 -g, LAT --optimize-O1:

• Previous optimized baseline: hot AOT 626-629.
• This clean branch: cold AOT 613; hot AOT 644.
• AOT was considered ready only after 5 files totaling 14,932,326 bytes remained unchanged for eight checks.
• Same-machine non-AVX reference: hot AOT 644.

The clean branch excludes jump-cache, profiler, AES/VAES, and AOT build-identity changes.

Validation / 验证

• Clean x86_64 LAT build passed (407/407 build steps).
• Clean i386 LAT build passed (407/407 build steps).
• meson test -C build64-pr --suite lat-pr-fast: 24 passed, 0 failed.
• avx128-shuffle.S, avx-sum3-pattern.S, repeat-add.S, and vcomis-jcc.S: native x86, LAT JIT, cold AOT, and hot AOT all passed with non-empty AOT files.

Checklist / 检查项

• I have read CONTRIBUTING.md. / 我已阅读 CONTRIBUTING.md。
• Every commit contains a DCO sign-off (git commit -s). / 每个提交都包含 DCO 签署。
• I have included relevant build, test, and performance results. / 我已提供相关构建、测试和性能结果。