Redesign SyclEvent class

[oleksandr-pavlyk]

dpctl.SyclEvent needs to closely follow SYCL interface.

C-API changes needed

• In C-API we need DPCTLEvent_Copy.

• We should either use e.wait_and_throw() in DPCTLEvent_Wait, or introduce DPCTLEvent_WaitAndThrow(ERef).

• Since sycl::event is a associated with a SYCL backend, we should support DPCTLEvent_GetBackend(ERef).

• sycl::event::get_wait_list() returning a vector of events, should be exposed via DPCTLEvent_GetWaitList(ERef).

• sycl::event::get_info<sycl::info::event::command_execution_status>() returning an enum class event_command_status : int { submitted, running, complete} should be exposed via DPCTLEvent_GetCommandExecutionStatus(ERef).

• sycl::event::get_profiling_info<sycl::info::event_profiling::command_submit>(), sycl::event::get_profiling_info<sycl::info::event_profiling::command_start>() and sycl::event::get_profiling_info<sycl::info::event_profiling::command_end>() returning uint64_t value of device timer corresponding to the logged time of referenced operations. These are blocking operations (i.e. wait on the event), and throw sycl::invalid_object_code if the queue that created the event did not have profiling enabled.

Python API changes needed

• dpctl.SyclEvent() should create object corresponding to sycl::event().
• dpctl.SyclEvent( _SyclEvent ) should copy event from _SyclEvent like other constructors do. _SyclEvent is the helper data owner responsible for calling DPCTLEvent_Delete. Do not think __dealloc__ needs to call DPCTLEvent_Wait().
• dpctl.SyclEvent( event_capsule ) where event_capsule is named "SyclEventRef" and stores pointer to sycl::event object should create a dpctl.SyclEvent carrying a copy of that sycl::event object.
• dpctl.SyclEvent.execution_status should return either an enum, or a string
• dpctl.SyclEvent.profiled_info.submit, dpctl.SyclEvent.profiled_info.start, dpctl.SyclEvent.profiled_info.end would return profiling data.
• dpctl.SyclEvent.get_backend() should return backend enum.
• I think dpctl.SyclEvent.wait() should call e.wait_and_throw().
• dpctl.SyclEvent.wait_list() should return a list of dpctl.SyclEvent instances
• Static method dpctl.SyclEvent.wait( event or a list of events ) should be implement for synchronously wait_and_throw on the given event or list of events.

N.B.: The functionality to combine multiple events into a single event perhaps belongs to dpctl.SyclQueue, where it would submit an empty single_task, and use depends_on the given list of events.

[oleksandr-pavlyk]

In Python one can not have static method wait alongside a regular wait method.

Therefore the static one is called wait_for(events).

Profiling info is queried using event.profiling_info_submit, etc.

Other than noted discrepancies, all of this was implemented.