Simplify execution graph via analysis of dependencies

[Alexandr-Konovalov]

If a dependent command has all dependencies that has newly added command, the dependent command can be excluded from the execution graph.

For in-order queue, all dependent command from current queue can be excluded from an execution graph.

In current implementation, case with single accessor is supported, but for
auto e = queue.submit([&](auto &h) {
sycl::accessor acc(DataBuf, h);
sycl::accessor acc1(DataBuf1, h, sycl::read_only);
h.parallel_for(range, ...);
});
called in loop a size of dependency graph is limited by LeafLimit.

[sergey-semenov]

Just to capture what we discussed here. Instead of adding a new condition where we send commands to cleanup, I think a better solution would be excluding such dependencies from leaves (where both the newest node and its dependency are leaves of the same memory object) and let the usual cleanup path handle it. That can be done by adding a redundant dependency to the graph, the same way we're keeping graph width under control by imposing a leaf limit on each memory object.

That approach would also cover more cases: it's not needed that every dependency of the old command is covered by the new one for us to be able to clean it up. We should only care about how many memory objects the old command is a leaf for, other dependencies shouldn't be relevant.