Add indexing support to SkipMissing

[nalimilan]

Define IndexStyle, eachindex, keys and getindex to match indices of wrapped array, skipping missing entries. This notably allows finding the index of elements in the wrapped array using findall, findnext, argmin/max and findmin/max.

Fixes #29305. Part of #30606.

FWIW, iterating over SkipMissing using eachindex and getindex rather than iterate is slower, but not that much (and I see no reason why the compiler couldn't improve since the involved code is quite simple) EDIT: with @inbounds annotations both solutions have the same performance (but the performance difference persists without it).

julia> function f(x)
           s = 0
           @inbounds for v in x
               s += v
           end
           s
       end
f (generic function with 1 method)

julia> function g(x)
           s = 0
           @inbounds for i in eachindex(x)
               s += x[i]
           end
           s
       end
g (generic function with 1 method)

julia> x = skipmissing(ifelse.(rand(10_000) .> 0.8, missing, rand(10_000)))
Base.SkipMissing{Array{Union{Missing, Float64},1}}(Union{Missing, Float64}[0.764807, 0.225138, missing, 0.920196, missing, 0.782529, 0.936112,0.0176555, 0.581311, 0.624459  …  0.175387, missing, 0.393418, 0.717713, 0.898142, 0.899193, 0.177139, missing, 0.56281, 0.479712])

julia> using BenchmarkTools

julia> @assert f(x) == g(x)

julia> @btime f(x);
  24.347 μs (1 allocation: 16 bytes)

julia> @btime g(x);
  24.521 μs (1 allocation: 16 bytes)

[StefanKarpinski]

I'm not sure that I'm the best person to review this. @mbauman, would you feel qualified to take a look at this and review it?

[nalimilan]

Given the small amount of code introduced, I think what's needed is mainly a decision regarding the proposal at #30606.

[mbauman]

This is great as is — just a few optional comments on docs that I noted as I was reading through.

[nalimilan]

Thanks! I've pushed a commit to try to improve the wording.

[nalimilan]

I've just realized that broadcast(f, skipmissing(x)) currently works and returns a vector of length count(!ismissing, x). This is somewhat inconsistent with the indexing support added in this PR: for example a[findall(skipmissing(a) .< 0)] is a bit misleading since it will always work but return values which don't correspond to anything. Probably not a big deal since it's quite convoluted and the only alternative is to throw an error, but maybe something to change in 2.0.

[mbauman]

Yup, that'll fit in nicely with my hope to do something about wider axis support beyond just AbstractUnitRanges. Linking to #24019 as a nice example of a type that's in-between associative and array. Also interesting that it's the first example that I've seen where the default broadcastable fallback does the wrong thing.