llama.cpp SYCL fix stops dual Arc RAM bloat

This looks like a driver-path trap, not a tuning problem: the model fit was never the issue, the allocation API was. If the fix upstreams cleanly, it removes a major blocker for practical multi-GPU Intel local inference.

• –`sycl::malloc_device()` appears to hit xe's DMA-buf/TTM path, so GPU allocations get mirrored into host RAM at allocation time
• –`zeMemAllocDevice()` avoids that mirror behavior, and the post claims SYCL kernels can still read the pointers without issues
• –The failure mode is nasty because kernel-side staging bypasses cgroup limits and looks like an application-level memory leak
• –The evidence points to a stability win more than a performance win: throughput stays flat while RAM usage drops from crash territory to roughly 10%
• –If reproducible broadly, this changes the hardware advice for dual Intel Arc setups from "buy more system RAM" to "use the right allocator"