Optional quadratic duration correction for dynamic bucketing sampler

[pzelasko]

I've also removed some deprecated options like specifying max frames/samples and related tests.

[pzelasko]

I gave it some runs, it seems to work alright, even though I still haven't calibrated it very well.

As a piece anegdotic evidence, I was able to adjust a training which had max_duration=250 into max_duration=400 with quadratic_duration=30 without OOM. So I almost doubled the duration, but the average duration in my dataset is about 5-6 seconds, which is far below the penalty that effectively doubles a cut duration. It's hard to give effective increase in throughput without measuring it carefully, but I think throwing around numbers like 20-30% is not unreasonable.

I'll merge it. If somebody uses it, it's important to remember that you need to adjust max_duration after setting quadratic_duration (and likely learning rate as well).

[danpovey]

That's great!!