remove unnecessary onload hack from mathjax macro

[minrk]

This is still up for discussion.

@damianavila can you provide an example where the hack is necessary? It seems it was added in a PR only meant to consolidate files without changing any behavior. I can find no report of the problems this claims to fix. I very much doubt that it is the right fix for whatever problem it addresses, so I would like to try to find the right solution.

[ivanov]

IIRC @matthew-brett recently wrestled with this hack, as well.

[damianavila]

The hack was originally introduced when IPython.nbconvert was an standalone app...
Surprisingly, I have tested your branch I do not see the problem anymore. The problem was essentially a bad rendering of the math symbols, and it had a random nature (a kind of heisenbug), only fixed by the hack of loading mathjax at the very end...
Anyway a lot of things changes, not only in IPython.nbconvert, but also in reveal.js, and the bug seems to be gone.
To conclude +1 to merge this one... because I can not replicate the problem anymore (which is a great notice, btw).
Thanks for start the discussion and "force" me to check this again 😉

[ellisonbg]

I don't recall why we put this hack in originally. But I am fine with merging this and seeing what happens. MathJax should work fine without this hack - for sure.

[minrk]

Merging away, we'll see if someone actually sees an issue from this. nbviewer has never had this hack, and it's by far the most used case for nbconvert + mathjax.