Fix getindex

[devmotion]

Fixes #107 and fixes #109 by using copyto! instead of unsafe_copyto!.

Additionally, I fixed type stability of getindex and support for empty ChainedVectors.

Currently on the main branch:

julia> x = ChainedVector([[1, 2], [3, 4]]);

julia> x[Int[]]
0-element ChainedVector{Int64, Vector{Int64}}

julia> x[2:1]
0-element ChainedVector{Int64, Vector{Int64}}

julia> x[[2,3]]
2-element Vector{Int64}:
 2
 3

julia> x[2:3]
2-element Vector{Int64}:
 2
 3

julia> x = ChainedVector([[]]);

julia> x[Int[]]
0-element ChainedVector{Any, Vector{Any}}

julia> x[2:1]
0-element ChainedVector{Any, Vector{Any}}

With this PR:

julia> x = ChainedVector([[1, 2], [3, 4]]);

julia> x[Int[]]
Int64[]

julia> x[2:1]
Int64[]

julia> x[[2,3]]
2-element Vector{Int64}:
 2
 3

julia> x[2:3]
2-element Vector{Int64}:
 2
 3

julia> x = ChainedVector([[]]);

julia> x[Int[]]
Any[]

julia> x[2:1]
Any[]

I don't see any performance regression with this PR in the benchmark in #105, on the contrary probably due to the type inference fixes the PR reduces the number of allocations:

Currently on the main branch:

julia> @b f($vinputs, $(1:95000))
10.083 μs (9 allocs: 372.125 KiB)

julia> @b g($vinputs, $(1:95000))
18.750 μs (8 allocs: 904.703 KiB)

With this PR:

julia> @b f($vinputs, $(1:95000))
8.875 μs (5 allocs: 372.016 KiB)

julia> @b g($vinputs, $(1:95000))
19.042 μs (8 allocs: 897.328 KiB)

[devmotion]

@quinnj would you be able to review the PR? We hit #107, the bug fixed in this PR, with a simple df[1:1522, :] subsetting operation of a DataFrame with ChainedVector columns that was created by CSV.read (Julia with multiple threads), and given that this seems a rather standard use of DataFrames + CSV I think it would be good to fix #107.

[quinnj]

Thanks @devmotion! This is great!