Support blocking for debugging dpcpp kernels

[yslan]

Hi,

CUDA and HIP provide the CUDA_LAUNCH_BLOCKING and HIP_LAUNCH_BLOCKING environment variables, which force synchronous kernel launches and make it easier to obtain a backtrace for a failing kernel. Since SYCL does not appear to offer a similar mechanism, would it be possible to introduce something like OCCA_LAUNCH_BLOCKING that forces a device–host synchronization after each kernel launch? This would at least help users identify the first failing kernel more quickly.

Something like manually calling device.finish() for each kernel launches?

Thanks!