[SYCL][ESIMD] Simplify the copy_to() and copy_from() implementation

[sarnex]

We can use block_load and block_store for these functions now in most cases, except when stateful mode is enabled and we're using accessors due to alignment and size requirements.

We also see better performance as seen in the perf test updates with fewer instructions.

The tests were written by Slava in an earlier version of this PR.