Run interpreted tests

@ezyang suggested in #2058 that a flag (such as `--interpreted`) be added to `cabal test` which would make Cabal run the test suites through the interpreter, rather than recompiling them. This will be faster in many instances because it avoids linking. We should _not_ do this for benchmarks, because interpreted code isn't optimized.

Start by locating the implementation of the `cabal test` command and how it distinguishes test suites from benchmarks. Trace how test suites are compiled and linked, then determine where an `--interpreted` option would be handled without affecting benchmarks. Done means interpreted test execution works through the command-line path and the existing test behavior remains unchanged.

Written by the indexing model from the issue text.