Parallel seeding

[thomas3494]

I replaced the inherently sequential seed_arr of complexity O(k) with an embarrassingly parallel algorithm of O(k log k) complexity. We should create a seed array of length p (or a small multiple) if we have p processors, so the parallel runtime
O(log(p)) will always be smaller than the sequential time O(p).

I also changed the seed_arr API to take any shape, rather than a number as argument. This way we can hide the ugly reshape workaround in Stdlib.

The function seed_arr(seed) computes [seed, seed + 1 * J, seed + 2 * J, ..., seed + k * J], where J is some large number. We can compute seed + K in constant time, if we have precomputed a magic number for that K. The idea is to compute a table of length 32, that has the magic numbers for 2^(j) * J, j = 0, ..., 31. We can then advance the state k * J times in log_2(k) steps by decomposing k into b_0 * 2^0 + b_1 * 2^1 + ... + b_(log_2(k)) 2^(log_2(k)).

I also push the sage code for computing the magic numbers, with references. This can be used to support other GF(2)-based PRNG's as well. The only math necessary to adjust this code, is the matrix form of the generator, all the ring-theory does not differ between generators.