Support non deterministic generator

[oleksandr-pavlyk]

1. Introduces support for non-deterministic hardware instruction RDRAND-based basic random number generator on relevant architectures.

In [1]: import mkl_random, numpy as np

In [2]: rs = mkl_random.RandomState(brng='NON_DETERMINISTIC')

In [3]: rs.randn(10)
Out[3]: 
array([ 1.69156146,  0.56649183,  0.26323201,  0.46801072, -0.97351586,
       -1.43365592, -0.73531088, -0.90349218,  0.23288929, -0.24443191])

Because it is hardware-based TRNG, reading pickled state does not produce the same randomness:

In [10]: import pickle

In [11]: rs = mkl_random.RandomState(brng='non_deterministic')

In [12]: rs2 = pickle.loads(pickle.dumps(rs))

In [13]: rs.randint(0, 2**32, size=5)
Out[13]: array([ 649482295, 3967010060, 2807993698, 1176520041, 2802939972])

In [14]: rs2.randint(0, 2**32, size=5)
Out[14]: array([  58520572, 2008182244, 2128388840, 2600218768, 2531220520])

compare to doing the same with pseudo-randomness generator:

In [15]: rs = mkl_random.RandomState(brng='SFMT19937')

In [16]: rs2 = pickle.loads(pickle.dumps(rs))

In [17]: rs.randint(0, 2**32, size=5)
Out[17]: array([3569610330,  417479194,  261680738, 3266235240, 3210854201])

In [18]: rs2.randint(0, 2**32, size=5)
Out[18]: array([3569610330,  417479194,  261680738, 3266235240, 3210854201])

2. Use Intel(R) Math Kernel Library viRngMultinomial directly to draw samples from multinomial distribution

In [19]: pvec = mkl_random.rand(100)

In [20]: pvec /= pvec.sum()

In [21]: pvec.sum()
Out[21]: 1.0

In [22]: %timeit mkl_random.multinomial(1000, pvec, size=10**5)
642 ms ± 1.02 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [23]: s = mkl_random.multinomial(1000, pvec, size=10**5)

In [24]: s[:, 0].mean()
Out[24]: 8.29136

In [25]: pvec[0] * 1000
Out[25]: 8.28811517092862