Whitespace not captured correctly from some PDFs

## Bug report form
### Steps to reproduce
1. Open https://via.hypothes.is/http://www.ps.uni-saarland.de/PapersOz/ProgrammingSysLab/ngc98.pdf#annotations:_8XCnOHJEeeAXlc1u38s-A
2. Select some text from the abstract
3. Click the highlight icon

### Expected behaviour
The text shown in the hypothesis panel on the right should match the text highlighted.

### Actual behaviour
The text shown in hypothesis does not include spaces from the original text, and is unreadable.
E.g. for https://hypothes.is/a/_8XCnOHJEeeAXlc1u38s-A
"Thereisnoglobalinformationnorglobaltime." in hypothesis
"There is no global information nor global time." in the original document

### Browser/system information
Firefox 57 on Linux.

### Additional details
Only happens with some PDFs, not sure why.