Should the default SEXP in r_vector be a typed vector of length 0 rather than R_NilValue?

[DavisVaughan]

i.e. to support things like #315 where you default construct the vector and immediately try and put attributes on it.

Would also simplify internal code that currently has to account for the possibility of R_NilValue as data_ quite a bit.

We'd have to write our own default constructor but I think that's totally possible?

---

A downside of this is that the default constructor would have to allocate and would no longer be noexcept or as "cheap", but I can't see that mattering much because you eventually allocate anyways?