Symbol extraction improvements: merging with scripts and font formats

[andrewhead]

See TODOs 2 and 11 in issue #192 on symbol segmentation.

The main issue this PR seeks to solve is one of precision extraction of complex symbols from MathML.

LaTeX symbols are detected by:

1. Extracting equations
2. Parsing those equations into MathML using KaTeX
3. Cleaning up the MathML equations
4. Visiting the elements in the MathML to find elements corresponding to symbols

The problem is that the KaTeX is sometimes a little messy. For instance the word ReLU is parsed into four symbols (<mi>R</mi><mi>e</mi><mi>L</mi><mi>U</mi>). This means that if we want ReLU to be detected as a single symbol, we need to merge together consecutive nodes into longer words.

That much has already been done. But the behavior missed a couple of corner cases. This PR addresses those corner cases. What it specifically supports is:

1. Merging the bases for script elements (e.g., turning the parse tree for word_i---<mi>w</mi><mi>o</mi><mi>r</mi><msub><mi>d</mi><mi>i</mi><msub> (where 'd' is the base)---into the simpler <msub><mi>word</mi><mi>i</mi></msub> (where 'word' is the base))
2. When all the characters in a row are part of a single formatted word (e.g., from the macro \mathcal{word}, making sure that the merged node <mi mathvariant='script'>word</mi> is given the style attribute mathvariant, and its character offsets include the macro (and not just 'word').