Factor out BenchmarkGroup

[LilithHafner]

BenchmarkGroup is a useful datastructure for organizing hierarchical data that need not be BenchmarkTools.jl specific. It is similarly well-suited to storing benchmark results, test results, statistical conformity results (e.g. for random samplers), performance figures for ML algorithms, etc. For example, if folks are using Chairmarks with AirspeedVelocity, they may want this type without the rest of BenchmarkTools.

Additionally, I suspect that this conceptual and technical separation between benchmarking and handling nested data would make it easier for folks to wrap their heads around how to use BenchmarkGroups. Is this something you're open to & do you think it's a good idea, @willow-ahrens?

I'd imagine the type would be a NestedDict, BenchmarkGroup would be an alias for NestedDict, NestedDicts.jl would live in the JuliaCI or the JuliaCollections organization (or maybe NestedDict would live in DataStructures.jl, but BenchmarkTools doesn't already depend on DataStructures, and I wouldn't want to add a new big dependency). This would not be a breaking change.

[willow-ahrens]

This might be helpful if the type were more specific to nested groups of experiments. For example, nested testsets might have more than key-value pairs (broken, skip, failfast, etc.). Some approaches support tagging, where more than one tag is attached to a testset so we can specify groups to run with a few related tags. I'd also like to one day redesign the BenchmarkResult type to be parametric with respect to perf results, but I have found less time than I'd like for maintenance these days.