`find_emph_char` returns 0 if the char specified is not found in the
current line, but this is also what happens when there's an empty
cell. This patch adds logic to work around this problem.
See uranusjr/macdown#321
When the `HOEDOWN_EXT_UNDERLINE` extension was enabled, underlined spans
would actually be passed verbatim to the output buffer. This was because
the active_char was _only_ set when the emphasis, double_emphasis, or
triple_emphasis handlers were registered. As a result, no active char
was found in the input buffer, so everything was passed through
verbatim.
This patch fixes this by also registering the `active_char` if the
underline handler is registered. I also added a simple regression test.
I personally don't use this extension, but I encountered this bug over
the course of writing bindings for Rust.
