LATX, fix: Avoid LL/SC lowering for locked i386 operations

[baibaidashixiong]

Summary / 变更说明

• Steam's i386 runtime on LoongArch received EFAULT from a private
  FUTEX_WAIT even though the guest futex word remained mapped. The root
  cause is not the futex syscall itself: an unaligned i386 LOCK operand can
  enter the LL/SC interpreter fallback, which temporarily mremaps the host
  page. A concurrent host futex wait can observe that short-lived unmapped
  window.
• Use software-lock lowering for all locked i386 instruction translators,
  including the previously separate LOCK CMPXCHG8B path. i386 uses one
  process-local lock rather than a 64-byte hash, because overlapping x86
  operands can start in different hash stripes. x86-64 keeps its existing
  lowering.
• Release the current i386 lock owner on every nonlocal CPU-loop exit, so a
  guest page fault cannot strand the lock. This is a root-cause repair: it
  removes the temporary page migration instead of retrying a failed futex.
• The single i386 regression combines unaligned LOCK INC, LOCK CMPXCHG8B,
  private FUTEX_WAIT and FUTEX_WAIT_BITSET, a cross-stripe overlap check,
  and a PROT_NONE fault-cleanup check. It runs the complete
  LATX_AOT=0/1 × LATX_KZT=0/1 matrix with an isolated temporary HOME for
  AOT cache files.
• The process-local lock covers the Steam/private-mapping case. Cross-process
  MAP_SHARED x86 LOCK atomicity needs a separate shared-lock design.

Validation / 验证

• ninja -C build32 latx-i386: passed on LoongArch64.
• ninja -C build64 latx-x86_64: passed on LoongArch64.
• Focused product test: all four LATX_AOT=0/1 × LATX_KZT=0/1 combinations
  passed. The AOT=1 passes share the temporary cache directory after its
  first creation.
• Product strace of the focused test observed roughly 100,000 futex calls
  for each KZT mode, with zero futex EFAULT results and zero mremap calls.
• meson test -C build32-tests --suite latx-integration --print-errorlogs:
  passed on the PR branch (5/5) and again after integration (6/6).
• git diff --check: passed. Independent review found no blocking issue;
  it verified the i386 LOCK paths, CMPXCHG8B, cross-stripe overlap, and
  nonlocal-exit cleanup.
• No performance result is claimed. The change addresses an exceptional
  correctness path rather than a throughput optimization.

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（git commit -s）。
• I have included relevant build or test results, or explained why they
  are not applicable. /
  我已提供相关构建或测试结果，或说明了不适用的原因。