Avoid some unnecessary strlen() calls in exception text formatting, while still keeping up the nice property of sharing const C message strings.

Revert "Do not compile Lexicon.py. It's really only used once and is much larger as a binary module than as a byte code file, but not slower to import."
Currently seems to be required when the scanner is compiled.
This reverts commit d66405c6.

Do not compile Lexicon.py. It's really only used once and is much larger as a binary module than as a byte code file, but not slower to import.

Include deprecation warning from CPython (3.3+) in __Pyx_PyNumber_IntOrLong() and avoid inlining error handling code.

Streamline __Pyx_PyNumber_IntOrLong() with direct slot calls instead of C-API function calls that only do what we already do on our side already.

Specialise coercion from bytes to char* (and friends) to avoid going through several type checks in the generic "__Pyx_PyObject_AsStringAndSize()" function.

Split uncommon case from unicode/bytes.endswith() implementation to suggest inlining the normal case.

Fix reference leak in the Py2.x __Pyx_ReleaseBuffer() implementation for types with an overridden __releasebuffer__ method.

Closes #1638.

Fix more tests: Correct the source character position of name nodes to improve error reporting (previously reported the character position *behind* the name).

Correct the source character position of name nodes to improve error reporting (previously reported the character position *behind* the name).

Closes #1824.

Repair "range(enum)" iteration in C++ which (rightfully) disallows arbitrary assignments between enum+int.

Revert "Re-enable failure checks for Py3.2 in travis as that version is correctly supported (again)."

This reverts commit 835cf4c1.

The travis test runs seem to be unreliable for some reason.