Fix bit order of RMAT Rectangular Generator to match expectation

[mfoerste4]

Previously, the decimal indices of rows/columns of the adjacency matrix did not align with the node-ids created by the algorithm. This PR fixes the bits set for each decision during the computation as described by the docstring.

FYI @tfeher

[tfeher]

Thanks Malte, this looks good! Just one additional request: it would be great to update the docstring of thetha to be explicit about which element of the array describes which bit level.

I would recommend to start with the main docstring, and replace

raft/cpp/include/raft/random/rmat_rectangular_generator.cuh

Lines 33 to 34 in 596d4b7

	* This is the most general of several overloads of `rmat_rectangular_gen`
	* in this file, and thus has the most detailed documentation.

with

* This function generates a random graph represented by a (sparse) adjacency matrix. As described
* in [1], to generate connections, we recursively subdivide the adjacency matrix into four
* equal-sized partitions, and distribute edges within these partitions with a unequal
* probabilities. The probabilities are described by numbers [a, b, c, d]. We chose the upper left
* partition with probability `a`. The chosen partition is again subdivided into four smaller
* partitions, and the procedure is repeated until we reach a single element (1 × 1 partition).
*
* We can prescribe different probability distribution at each iteariton. The `theta` array stores
* the probability values for each level.
*
* [1] "R-MAT: A Recursive Model for Graph Mining" Deepayan Chakrabarti, Yiping Zhan, Christos
* Faloutsos (2004) https://doi.org/10.1137/1.9781611972740.43

After this, we can clarify the layout of theta by replacing

raft/cpp/include/raft/random/rmat_rectangular_generator.cuh

Lines 52 to 56 in 596d4b7

	* @param[in] theta distribution of each quadrant at each level of resolution.
	* Since these are probabilities, each of the 2x2 matrices for
	* each level of the RMAT must sum to one. [on device]
	* [dim = max(r_scale, c_scale) x 2 x 2]. Of course, it is assumed
	* that each of the group of 2 x 2 numbers all sum up to 1.

with

* @param[in]  theta   array [on device] with the distribution of each quadrant at each level of
*                     resolution. theta = [a0, b0, c0, d0, a1, b1, c1, d1, ...], where 
*                     [a0, b0, c0, d0]  defines the probability at the finest level (2x2).
*                     The last four elements in the array describe the probability in the
*                     coarsest level (where matrix size = [2^r_scale, 2^c_scale]).
*                     Since these are probabilities, the four [a_i, b_i, c_i, d_i] values for
*                     each level of the RMAT must sum to one.
*                     [dim = max(r_scale, c_scale) x 2 x 2]. 

What do you think?

[mfoerste4]

@tfeher , thanks for the good doc-string suggestion, I applied it to the branch.

[tfeher]

Thanks Malte for the update, LGTM!

[tfeher]

/merge