Improve SPAN model convergence prediction ETA

[olegs]

At the moment ETA is falling down significantly during first iterations, because it takes coverage computation time into account.

[Mar 31, 2019 15:46:33] 1.20% (3/250), Elapsed time: 3 min 3 s, Throughput: 1 min/item, ETA: 4 h 11 min
[Mar 31, 2019 15:47:04] 2.00% (5/250), Elapsed time: 3 min 34 s, Throughput: 1 items/min, ETA: 2 h 55 min
[Mar 31, 2019 15:47:48] 3.20% (8/250), Elapsed time: 4 min 18 s, Throughput: 1 items/min, ETA: 2 h 10 min
[Mar 31, 2019 15:48:17] 4.00% (10/250), Elapsed time: 4 min 47 s, Throughput: 2 items/min, ETA: 1 h 55 min
[Mar 31, 2019 15:49:01] 5.20% (13/250), Elapsed time: 5 min 31 s, Throughput: 2 items/min, ETA: 1 h 40 min

I suggest explicitly compute coverage before model fitting.