[SYCL] optimize createSyclObjFromImpl calls to take rvalue-ref to shared_ptr

[lslusarczyk]

The optimization results in moving shared_pointer inside createSyclObjFromImpl instead of copying and thanks to it we save two atomic operations (see e.g. this SO thread).

I've applied it to all possible places in the code, leaving only these where copying is indeed needed (mostly for context_impl use).

Results summary

overhead over UR reduced by ~8% in scenarios using events. Other benchmarks also show visible improvements in many cases, including new pytorch multiqueue benchmarks which improved overall by 2.7%

Results Examples

The new result is expressed by dots on the right sides of plots.

old = 134.6, new = 132.8, UR baseline = 113, overhead over UR reduced by 8.3%

old = 140, new = 138.2, UR baseline = 118.1, overhead over UR reduced by 8.1%

old = 122.3, new = 121.3, UR baseline = 108.1, overhead over UR reduced by 7.0%

old time = 13.91, new time = 13.58, whole stack reduced by 2.4%

And finally new pytorch microbenchmarks:

old time = 1.81, new time = 1.76, L0 baseline = 1.44
whole stack reduced by 2.8%, overhead over L0 reduced by 13.5%