Execute graph using handler

[EwanC]

Change the queue::submit(command_graph<executable>) API for launching an executable graph to handler::exec_graph(command_graph<executable>). See Issue #21

Using the handler is more in-keeping with the existing SYCL API and allows the execution of graph to depend on an arbitrary event through handler::depends_on. This design makes it easier for users to write the following code without have to block in host for waits.

auto ev1 = q.submit([&](handler& cgh){
  cgh.exec_graph(g);
});

// dest is some input to graph `g`
auto ev2 = q.memcpy(dest, src, numBytes,  ev1);

auto ev3 = q.submit([&](handler& cgh){
  cgh.depends_on(ev2);
  cgh.exec_graph(g);
});

Queue shortcut functions are also included in the spec, as is the case in the core SYCL spec for other handler functionality.

This change also enables the explicit API to capture nested sub-graph executions, which is not currently possible in the explicit API but is possible in the record & replay API. See issue #23

For example, with this change a user can now do:

command_graph<executable> executable_graph;
auto node = recordable_graph.add([&](handler& cgh){
  cgh.exec_graph(executable_graph);
});

[Bensuo]

LGTM! Just a minor suggestion.