Reduce allocations

[rafaqz]

This PR reduces array allocations for some nice performance improvements. It makes missing_omit _maybe_missing_omit, which does nothing when none of the columns have Missing eltypes.

I suspect modelcols can be much further optimized by reallocating and using a loop rather than reduce(vcat, x). I may add that too - this the main performance bottleneck after this PR.

[ararslan]

If I recall, the fact that this makes a copy even if T doesn't allow Missing was intentional to ensure that every column gets copied and downstream users of this function can mutate the result without worrying about affecting the original data. If we don't care about that, we can do

Suggested change

	_missing_omit(x::AbstractVector{T}) where T = copyto!(similar(x, nonmissingtype(T)), x)
	_missing_omit(x::AbstractVector{T}) where {T>:Missing} = copyto!(similar(x, nonmissingtype(T)), x)
	_missing_omit(x::AbstractVector{T}) where {T} = x

without needing any other changes.