speed up compilation for functions with large numbers of temps (already shows for a couple of thousand, e.g. for large list literals)

Docstrings don't support f-strings in cpython, so they don't need to here either. The scanner
seems to be working.

Current status:
- All but three of the tests from CPython's test_fstring pass. Two of the remaining ones rely
  on exec and the third relies on a runtime NameError that occurs at compile time in Cython. These
  tests still need to be added to the Cython test suite.
- I'm not overly convinced that the refcounting is correct in the C code that's currently
  emitted.
- There's some room for optimization, such as removing empty strings from JoinedStrNode,
  concatenating adjacent raw strings at compile time, and using _PyUnicodeWriter.
- Currently the feature is available in both Python 2 and Python 3 mode. It mostly works in
  Python 2, except for the !a conversion char.

Not sure about the refcounting

Using _PyUnicode_Writer might be faster but seems more complicated to implement.

Still doesn't do concatenation of FormattedValueNodes correctly.

Parses f-strings into ExprNodes, but with no support for compiling them. Some
known bugs are marked as TODOs.

This partially revert the change introduced in
767fce81.

Turns out AttributeNodes always start with type error, which we
don't want to propagate in some places.

Add visibility kwarg to cython.declare.

Add declarations for C++11/14 smart pointer factories (make_shared & make_unique)

DOCS: Fix typo