DeepSpeed

6e0a597d - XPU: bind to torch's SYCL runtime instead of loading a second one

XPU: bind to torch's SYCL runtime instead of loading a second one Resolving prepare_for_device_copy through the libsycl.so development symlink dlopen'd a second SYCL runtime alongside the libsycl.so.8 torch already had loaded. Both ended up with their own driver state and the process aborted at interpreter shutdown with UR_RESULT_ERROR_UNINITIALIZED. Look the symbols up on the soname with RTLD_NOLOAD so we only ever bind to torch's instance. Signed-off-by: Ma, Guokai <guokai.ma@intel.com>