[SYCL] HostTask deadlock fix

[cperkinsintel]

Multiple worker threads submitting to an in-order queue can deadlock when the application holds a mutex across q.submit() calls whose host_tasks acquire that same mutex.

The solution is to drain the synchronous host-task dep waits once, at the top of GraphProcessor::waitForEvent, before calling enqueueCommand. A new collectPendingHostDepEvents walks Cmd's dep graph under the read lock and gathers every host-task event that would otherwise be waited on synchronously. Then we release the read lock, wait for them all, reacquire, and proceed with the normal enqueueCommand. Its internal waitForPreparedHostEvents calls now find events already signaled and return immediately, so the read lock is no longer held across any blocking host wait.