This makes it much easier to update the bounds.  Instead of coming up
with a suitable (min,max) pair, you just give e.g. (base, 10) to allow
10% deviation from the base figure, which can be pasted from the
error.

e.g. previously:

-      # expected value: 458700632 (amd64/Linux):
-      if_wordsize(64,
-          compiler_stats_num_field('bytes allocated', 440000000,
-                                                      480000000)),

now:

+      if_wordsize(64, # sample from amd64/Linux 15/2/2012
+          compiler_stats_range_field('bytes allocated', 360243576, 10)),

Note: use stats_range_field rather than stats_num_field.  I left
support for the old way for now so that we can do a gradual migration.

(next I suppose we should make it so that 'make accept' works for perf
tests, but that's for another day)

Motivation: we wanted to disable -dcore-lint for the perf tests

Otherwise the colon confuses make on Windows

The : in c:/foo/bar on Windows confuses make.

I've sent them upstream here:
    https://sourceforge.net/tracker/?func=detail&aid=3484051&group_id=163236&atid=826945

I'll send them to pretty upstream.

I've sent them to the upstream bytestring bug report addresses

On OS X, /tmp is a symlink to /private/tmp, so the test prints the wrong
output. I've changed it to use /dev instead, which is hopefully less
likely to be a symlink.

It was explicitly removing the program inbetween the two runs, but
I don't see why that should be necessary.

We now just sleep for 2 seconds, rather than using
    touch --date="now + 3 seconds"
which doesn't work on OS X. OS X does have "touch -A", but that doesn't
work on Linux.