[SYCL] Add implementation of host-interop-task and test.

[s-kanaev]

This patch is number two in series of patches for interop part of host task.
This patch introduces an API to enqueue host-task with interop_handle argument
See the proposal [1].

Depends on #1937

[1] https://github.com/codeplaysoftware/standards-proposals/blob/master/host_task/host_task.md

[s-kanaev]

@Ruyk, @StuartDAdams I bet you know how to improve testing here.

[Ruyk]

Some ideas:

• A test that uses OpenCL interop to copy data from buffer A to buffer B , by getting cl_mem objects and calling the clEnqueueBufferCopy. Then run a SYCL kernel that modifies the data in place for B, e.g. increment one, then copy back to buffer A. Run it on a loop, to ensure the dependencies and the reference counting of the objects is not leaked. We could easily do the CUDA variant even with a macro later on.
• Same as above, but performing each command group on a separate SYCL queue (on the same or different devices). This ensures the dependency tracking works well but also there is no accidental side effects on other queues.
• A test that does a clEnqueueWait inside the interop scope, for an event captured outside the command group. The OpenCl event can be set after the command group finishes. Must not deadlock according to implementation and proposal, sketch below:

cl_event userEvent = clCreateUserEvent(...)
q.submit([&](handler& ) {
   h.codeplay_host_task([=](interop_handler& ih) {
     clWaitForEvents(1, &userEvent); 
   }
});
clSetUserEventStatus(userEvent, CL_COMPLETE);
q.wait();

[nyalloc]

Good stuff. Got a few comments on code style and a few things I'd like to clarify, but happy to see this ball rolling.

[s-kanaev]

@Ruyk, @StuartDAdams ping

[Ruyk]

There are two really minor changes required for the CUDA backend to work, but they should not block this PR. It seems to work fine otherwise.

[Ruyk]

Are all the lit testing passing for you? host-task lit test sometimes deadlocks on my system,

[New LWP 30329]
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
0x00007f8f775c5f30 in cl::sycl::detail::Command::enqueue(cl::sycl::detail::EnqueueResultT&, cl::sycl::detail::BlockingT) ()
   from /home/ruyman/open-source/build/lib/libsycl.so.2
(gdb) up
#1  0x00007f8f775db425 in cl::sycl::detail::Scheduler::GraphProcessor::waitForEvent(std::shared_ptr<cl::sycl::detail::event_impl>) ()
   from /home/ruyman/open-source/build/lib/libsycl.so.2
(gdb)
#2  0x00007f8f775d735d in cl::sycl::detail::Scheduler::waitForRecordToFinish(cl::sycl::detail::MemObjRecord*) ()
   from /home/ruyman/open-source/build/lib/libsycl.so.2
(gdb)
#3  0x00007f8f775d9158 in cl::sycl::detail::Scheduler::removeMemoryObject(cl::sycl::detail::SYCLMemObjI*) ()
   from /home/ruyman/open-source/build/lib/libsycl.so.2
(gdb)
#4  0x00007f8f775e8ba8 in cl::sycl::detail::SYCLMemObjT::updateHostMemory() () from /home/ruyman/open-source/build/lib/libsycl.so.2
(gdb)
#5  0x000000000040a343 in cl::sycl::detail::buffer_impl::~buffer_impl() ()
(gdb)
#6  0x000000000040a319 in void __gnu_cxx::new_allocator<cl::sycl::detail::buffer_impl>::destroy<cl::sycl::detail::buffer_impl>(cl::sycl::detail::buffer_impl*) ()
(gdb) quit

[s-kanaev]

Are all the lit testing passing for you? host-task lit test sometimes deadlocks on my system

@Ruyk, please, try again with #1937 fix.

[Ruyk]

Yes, that seems to fix the problem.

[s-kanaev]

@StuartDAdams , @Ruyk the only latest changes are some stylistic ones. Please review and approve if you have no objections.

[Ruyk]

The CUDA variant I have seems to work locally (bearing two minor changes), so looks good!