Address memory issues on PDF grafting during OCR

[duckduckgrayduck]

If a detect a visible text layer in the document, we bypass merging in the newer text layer because of memory issues.

documentcloud/documentcloud/documents/models/document.py

Line 505 in 8535b6a

# merging when we need to flatten visible text causes excessive memory usage

We should try to fix this memory issue so that when OCR is called via an Add-On and the document has a visible text layer, we still graft the new text in.

[duckduckgrayduck]

Related to #336

[eyeseast]

Alternate approach: Don't graft text from Textract and use page positions JSON file instead. We need to leave a marker in the document metadata to load this.