In particular, the 3 values you most need to care about (framework
failures, unexpected passes, unexpected failures) are now together
in a block.

Uses look like
    stats_num_field('bytes allocated',
                    [(wordsize(32), 45648, 5),
                     (wordsize(64), 49400, 5)])
where the first matching triple will be used. e.g. we could override
the Win32 expected values with:
                    [(platform('i386-unknown-mingw32'), 41000, 5),
                     (wordsize(32),                     45648, 5),
                     (wordsize(64),                     49400, 5)])
with other 32-bit platforms falling through to the wordsize(32) case.

This makes it easier to give different values for different platforms,
while being sure that all platforms are covered.

This allows them to give framework failures.

I also had to change how setTestOpts works. Now, rather than applying
the options to the directory's "default options", it just stores the
options to be applied for each test (i.e. once we know the test name).

We now deepcopy the options for each test, which means that the test
helpers don't need to worry about sharing.

This will reduce the number of helper functions that we need

This gives hte list of tests with unexpected results, so that you can
easily run them again.

We now use *_range_field everywhere instead

dph: tests with dph-lifted-copy need to use '-fno-vectorisation-avoidance if scalar closures get too big

We were getting a
    Creating library file: dist\build\libHStest-1.0-ghc7.7.20130205.dll.a
message on stderr. We can't turn it off, so now the test just normalises
it away.

Otherwise, you get annoying failures if you run 'make fast stage=1'.

Signed-off-by: Austin Seipp <mad.one@gmail.com>

We now use pre_cmd rather than cmd_prefix wherever possible.
Also, pass "-s --no-print-directory" whenever we use pre_cmd.

Fixes the test on Windows