Math refactor

[ollpu]

There are three changes here. (1) could be done without the other two to fix the bug.

(2) & (3) might have some small performance implications in some cases. However, they should make it easier to disallow empty display math items $$$$, which I'm planning to do in a follow-up PR.

1. Populate math_delims fully instead of stopping at an invalid closing delimiter. Inline code already works like this, though it can't encounter invalid delimiters.

   Fixes a bug when parsing e.g. this: $x {$ $ } $x$

2. Disallow $$ matching a closing $ and then marching delimiters in make_math_span. Instead, retry scanning at the second position (really just a math_delims.find()).

   Before:

   $$x$
   ^--|    # make_math_span called with this span
   

   After:

   $$x$
   ^--|    # no match, move forward
    ^-|    # make_math_span called with this span
   
   

3. Remove the seen_first optimization from MathDelims. It doesn't work with the retry strategy (2).

[ollpu]

Ok, apparently I was mistaken about commonmark-hs disallowing empty display math items ($$$$). Let's not change that then.

How do we feel about changes 2 & 3? They simplify the overall logic and make_math_span quite a bit, but they're not really necessary at this point.

[notriddle]

Overall, I love how much simpler this is! Fix the silly nitpick in the test case, then r=me

[Martin1887]

Perfect, thanks!