Reduce memory usage at rest

[mstoeckl]

This PR partially addresses #14. See commit messages for details.

There is a new dependency on gio-unix-2.0, to provide the 20 necessary lines of code from g_unix_input_stream_new. This can be removed (at +38/-12 diff cost) if the code switches to using GdkPixbufLoader directly.

[emersion]

Overall this looks pretty good! I do wonder though whether loading the images from a FD is really beneficial. The inode matching allows us to de-duplicate symlinks (does it?), but comparing paths doesn't sound too bad and should account for 99% of the users. Maybe we should consider dropping the first commit?

[mstoeckl]

I do wonder though whether loading the images from a FD is really beneficial

The main benefits were all edge cases:

• Being able to deduplicate soft and hard links, as well as unusual paths, like a.png vs ./a.png
• Only needing to call open() once per image; instead of e.g. once per monitor hotplug. This ensured that the background could still be drawn if the image file was deleted or renamed before a new output was added.
• Slightly nicer error messages, distinguishing failure to open, vs e.g. bad image data.

Maybe we should consider dropping the first commit?

Dropped, and adjusted following commits. The FD loading can always be added in later, if anyone wants it.