Improve vbmap greedy approach performance.

It was noticed in MB-55746 that it took around 95 seconds for vbmaps be generated for 30 buckets. I think there a couple of improvement we can make.

1. We current generate 9 vbmaps and select the best map amongst those 9 maps for each bucket. The reason we generate 9 maps is historic and we could potentially investigate if we could reduce that to a smaller number.

2. Early in the vbmap greedy approach I decided to use maps to encode the RI and R matrices - but eventually it turned out we didn't need maps and encoding them as 2-D arrays would work too.

We could use https://github.com/google/pprof to see where it can be improved.

./vbmap --num-nodes 100 --num-slaves 10 --num-replicas 1 --num-vbuckets 1024 --output-format json --greedy -cpuprofile ./cpu.prof

go tool pprof -http=":8000" vbmap ./cpu.prof