implement three-argument similar method (from #94)

[piever]

Rebase + cleanup of #94. I've just noticed that before there was a bit of an ambiguity in the interaction with offset arrays.

In the previous version,

julia> using StructArrays, OffsetArrays

julia> s = StructArray(a=rand(4));

julia> similar(s, -1:2) # StructArray of OffsetArrays
4-element StructArray(OffsetArray(::Vector{Float64}, -1:2)) with eltype NamedTuple{(:a,), Tuple{Float64}} with indices -1:2:
 (a = 1.4080165e-315,)
 (a = 1.408020214e-315,)
 (a = 1.396197183e-315,)
 (a = 1.402099964e-315,)

julia> similar(typeof(s), -1:2) # OffsetArray of StructArrays
4-element OffsetArray(StructArray(::Vector{Float64}), -1:2) with eltype NamedTuple{(:a,), Tuple{Float64}} with indices -1:2:
 (a = 0.0,)
 (a = 1.39654137e-315,)
 (a = 1.476201036e-315,)
 (a = 0.0,)

whereas with this PR they both give OffsetArrays of StructArrays:

julia> using StructArrays, OffsetArrays

julia> s = StructArray(a=rand(4));

julia> similar(s, -1:2) # OffsetArrays of StructArrays
4-element OffsetArray(StructArray(::Vector{Float64}), -1:2) with eltype NamedTuple{(:a,), Tuple{Float64}} with indices -1:2:
 (a = 0.0,)
 (a = 1.39654137e-315,)
 (a = 1.396105643e-315,)
 (a = 0.0,)

julia> similar(typeof(s), -1:2) # OffsetArray of StructArrays
4-element OffsetArray(StructArray(::Vector{Float64}), -1:2) with eltype NamedTuple{(:a,), Tuple{Float64}} with indices -1:2:
 (a = 0.0,)
 (a = 1.39654137e-315,)
 (a = 1.396105643e-315,)
 (a = 0.0,)

This PR behavior is kind of the only one that is clean to achieve because of the way dispatch is set up (though getting both to return StructArray of OffsetArrays is probably also feasible but a bit more finicky). Is there a principled way to know which format similar(s::StructArray, rg::UnitRange) should return?

[piever]

UPDATE: the PR has been updated to intercept OffsetArrays overloads and consistently return StructArrays whose components are OffsetArrays.