ocl: improved determinism for small SMM sequences/batches

[hfp]

• Implemented cl_cache_dir (see https://github.com/intel/compute-runtime/blob/master/opencl/doc/FAQ.md#feature-cl_cache).
• Introduced opencl_libsmm_timer_t (opencl_libsmm_timer_device, opencl_libsmm_timer_device).
• Introduced environment variable OPENCL_LIBSMM_TIMER (device|host or 0|1).
• Made some code optional (ACC_OPENCL_CPPBIN, ACC_OPENCL_SEDBIN).
• Enable ACC_OPENCL_CACHEDIR in case of __DBCSR_ACC.
• Fixed regex (tune_multiply.sh).
• Adjusted some default values.
• Updated LIBXSMM for Daint-CI.
• Code cleanup.