Cloud Hypervisor told guests their L3 cache spanned every package

The host numbers are wider than the guest’s package layout, so the guest now gets a cache-sharing width clamped to one die. ## Notes 8a7928e806ee (Saravanan D, 2026-08-15). Fix guest L3 cache spanning packages. CPUID cache topology leaves 0x4 (Intel) and 0x8000001d (AMD) were left as the host reported them. Sharing widths used host APIC numbering. A host L3 wider than guest package spacing made the guest kernel merge every vCPU into one L3 domain. Worst without NUMA. With NUMA the kernel drops the LLC dependency but still reports the impossible topology. Clamp sharing IDs to the guest die and never widen past the host. Intel leaf 0x4 core IDs rewritten with the guest value. AMD CCX narrower than a die stays intact. Fixes #5742 and #8730. Landed on main. Latest release remains v53.0. No GHSA. Not e8dc35a7 VFIO DMA-gap, not 9336633a HTTP body limit.