Fetch before trying to atomically mark in Compressor

[no-defun-allowed]

Tracing is exceedingly slow on biojava if we don't test-and-test-and-set. My guess is that we get cache line ping-pong when many workers try to mark the same popular objects; but every policy I checked has some sort of test-and-test-and-set pattern, and this seems like a common sense thing to do.

Plotty for a detailed before-and-after, although both configurations have some other changes to Compressor which I will upstream later. But this change still speeds up upstream - here is before:

============================ MMTk Statistics Totals ============================
GC	time.other	time.stw	total-work.time.min	total-work.time.total	total-work.time.max	total-work.count	
14	8575.41	14920.23	inf	0.000	-inf	0	
Total time: 23495.63 ms
------------------------------ End MMTk Statistics -----------------------------

And after:

============================ MMTk Statistics Totals ============================
GC	time.other	time.stw	total-work.time.total	total-work.time.max	total-work.time.min	total-work.count	
14	8826.60	9548.96	0.000	-inf	inf	0	
Total time: 18375.56 ms
------------------------------ End MMTk Statistics -----------------------------