gpu: intel: ocl: add a size guard to stateless addressing flag
Adding guards for stateless addressing model:
- OpenCL GPU runtime only since SYCL does not allow allocating buffers bigger than 4GB
- Primitive uses buffers bigger than 4 GB
- Device must support 64bit addresses
Also, readding a check for max allowed allocation on SYCL runtime, since it leads to some tests failing instead of being skipped.
Fixes: MFDNN-14454 and MFDNN-14459.
make test perf-gpu
make test perf-gpu
make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb set test_scope=NIGHTLY
make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb set test_scope=NIGHTLY
make test perf-gpu
make test perf-gpu
make test perf-gpu
make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb set test_scope=NIGHTLY
make test perf-gpu
make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb