Fix header/footer removal for texts

The current header/footer identification scheme looks at the first/last logical paragraph on a page and attempts to match it against a regex to determine if it's a header/footer to be removed.

This has a number of failure modes:
- header & page number can be split into two separate blocks by the page layout analysis causing them to be in separate paragraphs
- page layout analysis can fail in ways that cause the header to not be the first logical paragraph (e.g. marginalia, noise, etc)
- scannos can introduce noise that causes the regex to fail

A more robust header/footer removal strategy would greatly improve the quality of epubs, particularly for text-to-speech where they're very distracting. Some scoring attributes might be things like physical page position, all caps, leading/trailing number patching pageno from scandata, edit distance from inferred header template, etc