AFAIK, the python re module does not have a C language component. It is a pure python regular expression implementation.
V8 is a notable exception. They use their code generation pipeline to JIT compile regexps to machine code, making it the fastest regexp engine around bar none (I think).
Libraries don't appear in the Comp.Lang.Benchmark Shootout, so CL doesn't do so well.
Of course, the downside is that formal grammars are often a bad fit for custom languages of the web, which tend to have incomplete/undecideable/context-dependent grammars. This would work for Markdown, though, since it does have a parseable grammar.
Using a real parser could probably cut half the code. Been meaning to make my own Markdown parser, everyone else's kind of sucks :-)
That's not an easy question to answer.
If you were to do a faithful port i.e. using using the same techniques as C/Go code, it would be very close. upskirt uses a traditional lexing/top-down-parsing approach. There's nothing in Python syntax that makes writing such code more compact than in Go. It's a bit tedious to write but the benefit is that (in C/Go) it gives the best speed because it minimizes the number of times each source character is looked at.
If you were to use a different approach e.g. brute-forcing your way through the text several times with regexpes, which is the most popular way of doing that in dynamic languages, the code clocks at ~2 thousand lines (like the implementation I use for my home-grown blog system https://github.com/kjk/web-blog/blob/master/markdown2.py)
If you were to use this technique in Go, the code would probably end up smaller than the manual lexing/parsing approach used in upskirt, but it would be significantly slower.
Interestingly, upskirt approach would probably be slower in Python than the regexp approach, because regexpes are heavily optimized C code and looking at individual characters of the string in Python code isn't particularly fast due to python interpretation overhead.