I think different Linux setups generate different output for these tests,
so we just ignore the output on Linux now.

This removes some duplication from the testsuite driver

When handling a test says reqlib('mtl') we cache the result, so we only
need to run ghc-pkg once no matter how many tests say it. This makes
quite a difference, especially on Windows.

Also, clean up the way we find the output file. From the comment:

# Finding the sample output.  The filename is of the form
#
#   <test>.stdout[-<compiler>][-<version>][-<wordsize>][-<platform>]
#
# and we pick the most specific version available.  The <version> is
# the major version of the compiler (e.g. 6.8.2 would be "6.8").  For
# more fine-grained control use if_compiler_lt().

I'll update the wiki too.

Avoids the need to write composes()

Now instead of saying, e.g.
    namebase_if_compiler_lt('ghc','6.9', 'foo-6.8')
you say
    if_compiler_lt('ghc','6.9', namebase('foo-6.8'))

pointed out by Claus Reinke

Allows us to use a different name base for foo.stderr for old compilers, e.g.
test('tc141',
     namebase_if_compiler_lt('ghc', '6.9', 'tc141-6.8'),
     compile_fail,
     [''])

somehow I forgot to add this yesterday, I'm mystified about where it went..

This is better than a top-level clean() when using threads

spawnlp doesn't exist in Windows, so this avoids using it.

Fixes a problem where /dev/null doesn't exist on mingw.

os.spawnv is broken on non-cygwin Windows (looks like it goes via a shell).

The user001 test in the unix package can fail when stdin comes from a
file.

Fixes some framework failures.

When working on a new foo extension, you can now put your tests in the
testsuite, set ProjectTags=-foo in mk/build.mk and skip_unless_tag('foo')
in the tests.