Allow mix of Symbol and Pair in join

[nilshg]

Currently we can do:

df1 = DataFrame(a = [1,2], b = ["a","b"], c = rand(2))
df2 = DataFrame(a = [1,2], B = ["a","b"], d = rand(2))

join(df1, df2, on = [:a => :a, :b => :B]) 
join(df1, df2, on = [:a]) 
join(df1, df2, on = [:b => :B], makeunique = true)

but it's not possible to do:

join(df1, df2, on = [:a, :b => :B])

This would be very convenient for cases where one has a large number of columns to match on, with only one or two that don't match and require a Pair.

[bkamins]

Makes sense and it should not be that hard to do.
Probably we should have something like internal normalize_on function that rewrites any on argument originally passed to be a Vector{Pair}, so that the later code can safely assume only one form (and this operation should be cheap relative to the other things join does).
Then in normalize_on we could make all the checks that the passed argument is valid.

Would you want to make a PR (if you do not have time I will have a look at this).

The only thing we have to be careful about when implementing it is that on master we allow joining more than two data frames in a single call (in which case we do not allow Pairs).

[bkamins]

See #2006 for a fix