[SYCL] Always store last event (for ioq) if scheduler was not bypassed

[igchor]

This fixes an issue for cases where the kernel depends on events coming from the scheduler (host task events or unenqueued commands). The logic in submit_impl for calculating noLastEventPath did not take into account all dependencies which could result in skipping storing last even if handler.finalize() ended up submitting work to the scheduler.

I also restored the lock on the submit path. It is needed to store the last event in case handler.finalize() returns an event coming from the scheduler.

[JackAKirk]

This looks like it has unintended changes to image.cpp in UR?

[igchor]

@intel/llvm-gatekeepers this is ready to be merge. The failure is related to #18932 and wrong exclusion check in lit.local.cfg - fix is here: #19061