The matplotlib backend is set to agg in the function _import_matplotlib(), which is called both (1) when resetting matplotlib for each example and (2) when scraping each code block for matplotlib plots (matplotlib_scraper()). The forcing of the backend to agg in the latter case means that an example is unable to make use of non-agg backends.

My particular use case is that I want to use mplcairo in an example to show how to use its blending operators, e.g., the example begins:

import matplotlib
matplotlib.use("module://mplcairo.base")
...

With the current code, the matplotlib backend is immediately set back to agg after scraping the first code block, and so the subsequent code blocks (and the eventual plot) do not work properly.

This PR replaces the call of _import_matplotlib() in matplotlib_scraper() with direct imports of matplotlib and matplotlib.pyplot, so the backend will not be set to agg when scraping for matplotlib plots. Recall that the backend will still continue to be set to agg when resetting matplotlib between examples. So, the result of this PR is that an explicit call to change the backend within an example will now persist through the rest of that example, but otherwise the backend will be agg.