New PDF watermark feature, rewritten text output handling and several bugfixes

Today's update (631f59d) brings with it the following changes:

New features

A new feature was added, making it possible to add a unique text to each downloaded PDF file. This feature is enabled by adding add_watermark=1 to an object's custom_init variables. PDFs available to anonymous users will never have a watermark added.

Completely refactor code used to generate HTTP responses for text content because of inconsistent text encoding semantics.
Centralize the generation of the Content-Disposition HTTP header to ensure standard compliance.
Improved the error message if no content was generated from a request handler.
Added flash error message if script objects throw exceptions. Also removed old method of conveying errors.
Default content type used by textual content in read request handler changed to be text/plain instead of application/octet-stream.

The thumbnail request handler would sometimes return the wrong Content-Length if client had already cached content.
Fixed internal server error if SVG with broken XML markup was displayed. Now the correct error message is shown.
Fixed several situations where we could be causing mojibake (garbled text) because of broken text encoding:
- Object type link with non-ASCII text in URL.
- Objects with use_resolve_path=1 in custom_init and a non-ASCII file name.