tests: make a stab at approximating wall-clock times
Vaguely empirical observations:
* ".py" tests are about an order of magnitude faster than ".t" tests
* dividing size by 1000 gives an approximation to wall-clock
run time (in seconds) that is not completely ridiculous.