Improve performance of exporting variable-width strings to numpy StringDType

[kylebarron]

The only way to create a string-typed numpy array from Python is from a list of Python string objects. This is very slow because you first need to copy your data into Python string objects.

(There are some low-level unsafe numpy C APIs, but I don't want to be mucking around manually in the numpy C API.)

We currently export variable-width string data first to a list of PyString and then to the numpy array constructor.

Hopefully the Rust-numpy crate will add safe APIs for passing a vector of String or &str to numpy from Rust.

[kylebarron]

I'll link PyO3/rust-numpy#505, which I created last year for reading variable-width string data, but I don't think there's a separate issue for writing yet.