Improved InitSampling function speed by 2.12 times

[RukhovichIV]

This PR is connected with #6411
Discarding elements from generators takes up most of the working time in InitSampling.
Since stdlibc++ doesn't have any random engines with o(n) complexity (little-o), we only can optimize the number of discarded elements.
std::bernoulli_distribution requires 64-bit input, so in previous version we had to discard twice as much elements as now.
This little optimization gives us ~2.12 speed up of InitSampling time, which translates into up to 16% speed up of the whole training time when subsampling < 1
The quality remains the same:

Mortgage dataset
version	training time	RMSE
original	35.479	0.009271
optimized	30.538	0.009262

Santander dataset
version	training time, s	InitData time, s	init improvement, %	Log Loss
original	249.196	28.784	0.00	0.16607
optimized	239.298	17.173	67.61	0.16610
no discard	224.347	11.221	156.51	0.16613

Higgs dataset
version	training time, s	InitData time, s	init improvement, %	Log Loss
original	34.235	14.373	0.00	0.09339
optimized	29.201	8.316	72.84	0.09409