test discovery: yield results progressively

[bukzor]

saves 20% wall time and 20% memory

In yelp main:

BEFORE:
69.66user 3.01system 1:21.89elapsed 88%CPU (0avgtext+0avgdata 5,806,912maxresident)k
224inputs+88outputs (1major+1,944,570minor)pagefaults 0swaps

69.84user 4.73system 1:23.84elapsed 88%CPU (0avgtext+0avgdata 5,801,632maxresident)k
0inputs+56outputs (1major+4,033,266minor)pagefaults 0swaps

68.77user 2.54system 1:20.60elapsed 88%CPU (0avgtext+0avgdata 5,938,432maxresident)k
0inputs+48outputs (1major+1,590,598minor)pagefaults 0swaps

AFTER:

64.93user 3.11system 1:08.89elapsed 98%CPU (0avgtext+0avgdata 4,779,600maxresident)k
0inputs+56outputs (3major+1,784,920minor)pagefaults 0swaps

63.14user 3.57system 1:07.30elapsed 99%CPU (0avgtext+0avgdata 4,779,792maxresident)k
0inputs+344outputs (1major+1,762,305minor)pagefaults 0swaps

64.71user 3.50system 1:08.77elapsed 99%CPU (0avgtext+0avgdata 4,779,600maxresident)k
8inputs+48outputs (1major+2,296,663minor)pagefaults 0swaps

Incidental changes:

• Removed the test_count reporting call back, which was used nowhere
• I believe we don't care about py26 or py33 anymore