It works well overall but there are a few issues:
- When vectors have different types, `promote_typejoin` is used (via `broadcast`)
  to choose element type. `promote_type` would be more appropriate but there is
  no mechanism to do this in Base.
- When skipping missing values, inference isn't able to realize that the result
  cannot be `missing`. This problem is fixed if I use a positional argument rather
  than a keyword argument for `skipmissing` but it's far from ideal for users.
- Since `eachrow(df::DataFrame)` returns a `DataFrameRows` objects which is also
  a table, `pairwise(cor, eachrow(df))` still computes correlation between columns
  rather than between rows. And anyway `DataFrameRow` is not accepted by `cor`
  since it's not an `AbstractVector`. One needs something like
  `(Vector(r) for r in eachrow(df))` to work around this limitation.
- Since Tables.jl objects can be of any type, we must have single method that
  performs dispatch internally by calling `Tables.istable`. Even `AbstractVector`
  inputs can be either vectors of vectors or row-oriented or column-oriented tables.
  This means the method that returns a named array and the method that returns a
  plain matrix have to live in the same package (which has to depend on NamedArrays).