reduce C code overhead for C++ std::string -> PyObject conversion helpers by implementing them directly in C++ and not Cython

--HG--
extra : transplant_source : %3E%C7%7C%E0I%E9%84%03%8A%EF%F0z%84%20d%E9%C1%12%AF%3E

DOC initialize string with bytes

As-is the example doesn't compile.

fix: not setting module reference to NULL in module init function fails to report exception on return

docs: mention that static typing with Py3 signature annotations requires the "annotation_typing=True" directive

Fix crash when an exception occurs during module init and the stale module gets used later on (after import failure).
This is fixed by no longer cleaning up the global references to module object and module dict, which obviously has the drawback of leaking memory.

provide "Cython.Compiler.Main.Version" to keep supporting old PyTables versions that import it from there

--HG--
extra : transplant_source : %B1%BC%5C%CD%A6%EEmr4B%0F%AF%1C%E0yq9%EA%ADX

avoid redundant type pointer comparison in "__Pyx_PyBaseString_Check()" as Py2.6+ already has fast builtin type checks

try not to write absolute source file paths into C file position comments and use paths relative to build directory instead

allow subtypes in right operand to override left side parent class behaviour for "@" operator (Python tries "__rmatmul__()" first in this case), generally clean up code duplication while at it

make libcython.py Py2/3 compatible to allows its immediate usage from both versions (and clean it up a little)

reduce excessive inlining in calling code a bit by moving some fallback code into a non-inline function

allow useless trailing semicolons in some places where people tend to put them, but warn about them (trac ticket 832)