The main elements introduced by this PR are:

1. The @xfail decorator to mark a test as an expected sanity failure.
2. The xfail() builtin to mark reference tuples as expected performance failures.

Two new test states are introduced which are also logged: XFAIL for expected failures and XPASS for unexpected passes.

Design-wise the concept is the same as the one explained in #2378.

Implementation-wise there are two major changes:

1. The execution policies code is refactored by eliminating code duplication between the serial and asynchronous policies.
2. A new utility function reference_bounds() is introduced which is used to validate and calculate the absolute lower/upper bounds. This used to be internal to assert_reference() and now it's made public (maybe the docs there need a slight reorganization).
3. The treatment of performance failures is rewritten as now we have to handle much more complex cases, were the performance variables of a test can simultaneously be pass, fail, xpass or xfail.

For more details on how feature works macroscopically check the tutorial and the reference docs.

Closes #2378.