Revisit DAG benchmarks

This overhauls the DAG benchmarks to address some long standing issues:

- Instead of measuring the combined time taken to generate a graph and
  all the graphs that come before it (i.e., TDAG -> CDAG -> IDAG), only
  measure the generation of the graph being benchmarked.
- Don't measure benchmark context creation / destruction.
- Fix IDAG benchmarks creating host tasks instead of device tasks (i.e.,
  number of devices and p2p copy support had no effect).
- Fix scheduler benchmarks comparing against single-threaded TDAG + CDAG
  generation, instead of all three graphs.
- Remove outdated "throttled submission" scheduler benchmarks.

Author: psalz

include/utils.h

@@ -19,6 +19,14 @@
 #define CELERITY_DETAIL_UTILS_CAT_2(a, b) a##b
 #define CELERITY_DETAIL_UTILS_CAT(a, b) CELERITY_DETAIL_UTILS_CAT_2(a, b)
 
+#define CELERITY_DETAIL_UTILS_NON_COPYABLE(classname)                                                                                                          \
+	classname(const classname&) = delete;                                                                                                                      \
+	classname& operator=(const classname&) = delete
+
+#define CELERITY_DETAIL_UTILS_NON_MOVABLE(classname)                                                                                                           \
+	classname(classname&&) = delete;                                                                                                                           \
+	classname& operator=(classname&&) = delete
+
 namespace celerity::detail::utils {
 
 /// Like std::move, but move-constructs the result so it does not reference the argument after returning.