Reduce RFENCE IPI cache flush scope

[jserv]

This optimizes RFENCE.VMA to use range-based cache invalidation instead of unconditionally flushing all MMU caches. It adds mmu_invalidate_range that selectively invalidates only cache entries within the specified virtual address range, reducing cache flushes by 75-100% for single-page operations.

SBI compliance: size==0 and size==-1 trigger full flush per specification.

---

Summary by cubic

Scope RFENCE.VMA cache invalidation to the requested virtual address range to reduce unnecessary flushes and speed up single-page operations. Full flush still occurs when size is 0 or -1 per SBI, reducing flushes by 75–100% on single-page cases.

• New Features
  • Added mmu_invalidate_range(hart, start_addr, size) and used it for RFENCE.VMA/VMA.ASID with hart mask selection.
  • Computed end of range with 64‑bit arithmetic and clamped to RV32 to avoid overflow.

^{Written for commit fb11de8. Summary will update automatically on new commits.}