Improve copy of SubDataFrame and DataFrameRow

[bkamins]

@nalimilan This is a follow up to #1534, and should be merged if the other is merged (I separate it as we have enough of changes there).

What it does:

• change copy for SubDataFrame to ensure that it returns a DataFrame in the future (this is the same functionality as we have but with a syntax taking into account future changes to getindex that are currently in deprecation period)
• introduce copy for DataFrameRow

Actually the second point is crucial as it is a simple change that is consistent with #1534 logic but shows one important thing. Should returned NamedTuple have minimal (narrowest) second parametric argument (type specification). This is what is implemented now.
Or it should have a wide type specification consistent with eltype of DataFrame column.

The difference is that either:

• narrow type specification will be possibly faster when NamedTuple is passed to a function and it might be a bit faster to create (not 100% sure though);
• wider type specification will mean that each NamedTuple created from a row of a given DataFrame will have the same concrete type.

I am not fully clear what is better (I have implemented first, but I have the feeling that maybe the second approach is better).

[nalimilan]

Thanks. I'd be inclined to use wider types instead, with the idea that it preserves more information and it makes things more "type-stable" in the sense that the type does not depend on the particular row (but of course inference doesn't know the columns types of a DataFrameRow). This is useful if you need to know what type to expect from remaining rows when allocating a column. It's possible that the wider type would be slightly less efficient (e.g. Union{T,Missing} rather than T), but taking a row in isolation doesn't make much sense.

Cc: @quinnj

[nalimilan]

See discussion at JuliaData/Tables.jl#28.

[bkamins]

@nalimilan
Given the decision in #1534 what copy on DataFrameRow should produce? A DataFrame or something else?

Also:

Base.convert(::Type{Array}, r::DataFrameRow) = convert(Array, parent(r)[row(r),:])

method seems strange given our current approach. Maybe we should deprecate it and only provide support for Vector only and then also add Vector(dfr) and Vector{T}(dfr) methods?

[nalimilan]

Maybe a NamedTuple would be better, as it is more similar to a DataFrameRow than a DataFrame (e.g. indexing gives a scalar).

The convert methods indeed need adjusting, but I think we'd better keep them as there's no reason to support only the constructor.

[bkamins]

OK.

With the convert methods the problem is that only convert(Array, dfr) now works and it produces a matrix not a vector. In particular convert(Vector, dfr) or convert(Matrix, dfr) currently fail.

[nalimilan]

OK. I guess we should at least support convert(Vector, dfr); it's a bit weird to see DataFrameRows as matrices.

Does this PR need anything more?

[bkamins]

I have added a doc string and it should be good.
I will handle conversions in a separate PR.