Improve memory footprint and performance of graph creation

[VibhuJawa]

This PR improves the memory footprint in graph creation and resolves story: #3439

The metric we are concerned about is input_to_peak_ratio which is currently 12.1x. Ideally this should be in 2x range.

Metric Meaning: How much memory expands for an input graph of size x. So if you want to analyze a graph of size x, we need 12x GPU ram currently.

Memory overhead ratio reduction:

• 12x to 6.14x for undirected (Reverse edges added)
• 5.9x to 3.07x for directed graphs (No reverse edges)

Time Reduction:

• 4.2x for scale 24
• 1.67x for scale 23

Comparison Table

scale	num_input_edges	directed	renumber	input_memory_per_worker	PR peak_allocation_across_workers	Main: peak_allocation_across_workers	Main: input_to_peak_ratio	PR input_to_peak_ratio
23	134217728	TRUE	TRUE	2.0GB	8.1GB	11.8GB	5.91	4.07
23	134217728	TRUE	FALSE	2.0GB	6.1GB	11.8GB	5.91	3.07
23	134217728	FALSE	TRUE	2.0GB	12.2GB	23.3GB	11.67	6.08
23	134217728	FALSE	FALSE	2.0GB	12.2GB	23.3GB	11.67	6.08
24	268435456	TRUE	TRUE	4.0GB	12.3GB	23.8GB	5.94	3.08
24	268435456	TRUE	FALSE	4.0GB	12.3GB	23.8GB	5.94	3.08
24	268435456	FALSE	TRUE	4.0GB	24.4GB	29.4GB	7.35	6.11
24	268435456	FALSE	FALSE	4.0GB	24.4GB	29.4GB	7.35	6.11

Graph Creation:

Scale	PR Time	Main Time	Speed Up
23	2.4	4.1	1.68
23	2.4	4.1	1.68
23	3.6	8.0	2.22
23	3.6	8.0	2.24
24	4.4	9.3	2.12
24	4.4	9.3	2.12
24	7.9	38.7	4.89
24	7.9	37.9	4.81
25	9.6	41.0	4.26
25	9.6	40.8	4.25

CC: @rlratzel

[alexbarghi-nv]

I think this PR may have some bugs, you may want to run the MG tests before finalizing it

[alexbarghi-nv]

I ran into issues trying to use this with the latest changes to sampling and edge properties

[VibhuJawa]

I think this PR may have some bugs, you may want to run the MG tests before finalizing it

@alexbarghi-nv , Can you try again. i think i have resolved all the bugs from my side and testing on multiple GPU configurations.

[alexbarghi-nv]

👍

[alexbarghi-nv]

/merge