Convert Python objects to C objects for further speedups

[sylvinus]

It is already done for nesting_limit. Each of the options should be transformed into C objects during init so that _traverse_node() uses as few Python objects as possible.

The gumbocy.html file generated after make cythonize is useful for seeing which lines use Python objects.

What would be the most efficient C type for lookups, to replace the Python sets like attributes_whitelist?

[sylvinus]

Most of the options have been converted to C variables.

There are probably some more optimizations left in the parsing of CSS class names (split in C instead of using Python's re?), but we should do more profiling first to see where the real bottlenecks are.

From my tests, >80% of the time is usually spent in gumbo.parse, not sure what we can do about it but look upstream for the largest speedups.

[sylvinus]

This one is also a good candidate for micro-optimization: 8e864c8#diff-51db9a1af8644d65b7f79981d2b0a7c2R62

[sylvinus]

A huge general speedup was gained thanks to #8, but it also re-introduced a lot of Python objects in the code.

There are a bunch of places where we go through Python strings for instance just to lowercase them. The attribute values are also stored as a Python dict, but a C++ map would probably be much faster (mostly because it would keep all its values as char*).