Make AxisKeys play nicely with AcceleratedArrays

Due to how lookups are handled in AK, it seems like in many cases the acceleration of AcceleratedArrays isn't used. For example:

using Dates
using AcceleratedArrays
using AxisKeys

dates = accelerate(Date(2018, 1, 1):Day(1):Date(2019, 2, 3), UniqueSortIndex)

sd = Date(2018, 3, 3)
ed = Date(2018, 8, 4)

i = findall(in(AcceleratedArrays.Interval(sd, ed)), dates) 
println(i)
@assert isa(i, StepRange) # nice, just 2 searchsorteds happened

A = KeyedArray(rand(length(dates)); date = dates)

# Not happy: AxisKeys is blind to AcceleratedArrays.Interval, crashes
# A(AcceleratedArrays.Interval(sd, ed), :)

# Works, but inefficient - the view created is not a range, but a Vector of Ints, since it just does Interval inclusion in a loop:
# https://github.com/mcabbott/AxisKeys.jl/blob/master/src/selectors.jl#L6
A(AxisKeys.Interval(sd, ed), :)

Could the lookup logic in AK be abstracted a bit, to allow other libraries to implement efficient axis lookup for their types? I think findall function is probably the correct interface, for example: https://github.com/kcajf/AcceleratedArrays.jl/blob/master/src/UniqueSortIndex.jl#L135. The currently implemented lookup logic could be reimplemented in terms of findall too, as a fallback.

Start with the lookup logic in src/selectors.jl and compare it with AcceleratedArrays' findall implementation in UniqueSortIndex.jl. Define an abstraction that lets AcceleratedArrays.Interval use efficient lookup while preserving the existing fallback behavior, then verify the interval selection example returns an efficient range.