llama.cpp issue #28134: SYCL backend aborts at model load when the device memory query fails

Lunar Lake Arc 140V iGPU on Ubuntu 26.04 with libze1 1.28.2 / Level Zero driver 26.05.37020.3 / oneAPI 2026.1: GGML_ABORT in ggml_backend_sycl_device_get_memory regardless of GGML_SYCL_GET_MEM_API=0/1 or ZES_ENABLE_SYSMAN=1. Also reproduces on a discrete Arc Pro B50 per comment LIXEREQ, who adds that b10698 works and b10729 does not. Maintainer arthw confirms GGML_SYCL_GET_MEM_API=1 rolls back to the legacy sysconf(_SC_PAGE_SIZE) path, which is Linux-only.