Slicing with larger key vectors is slow

Thanks for this great package! I find myself regularly slicing large KeyedArray matrices with large vectors of string keys (about 10% of the matrix). This is unfortunately currently slow:

using Random, AxisKeys, BenchmarkTools
A = KeyedArray(zeros(100000, 100), sid=["S$i" for i in 1:100000], oid=["O$i" for i in 1:100])
sub_sids = rand(axiskeys(A, 1), 10000)
A_sub_slow = @btime A[Key(sub_sids), :] 
# run time: 6.156 s

On my real data it can take minutes, which is why I regularly find myself using an indexin workaround:

A_sub_fast = @btime begin
    sub_sids_idx = indexin(sub_sids, axiskeys(A, 1))
    A[sub_sids_idx, :]
end 
# run time: 26.203 ms

@assert A_sub_slow == A_sub_fast

The slow method becomes much faster when slicing the matrix at the beginning (sub_sids = axiskeys(A, 1)[1:10000],
255.530 ms) and much slower when slicing at the end (sub_sids = axiskeys(A, 1)[end-10000:end],
18.904 s). The fast method is faster in all scenarios: 12.335 ms (slice beginning), 15.037 ms (slice end).

Only when performing small slices (100 elements) at the beginning can I see advantages of the default method (36.438 μs vs 77.185 μs). I may be missing something, but could the current method perhaps make use of indexin internally?

Start by locating the KeyedArray indexing path handling A[Key(sub_sids), :] and inspect how it resolves string keys. Reproduce the issue with the provided benchmark, compare it with the indexin workaround, and consider the work complete when large key-vector slicing matches the workaround's results without the reported slowdown.