gstdec: Avoid leaking memory when reading audio data

[ssssam]

We were reading audio data with the Gst.Buffer.extract_dup() method.
This allocates new memory using g_malloc() and returns it to the caller.
The memory needs to be freed with g_free(), however the PyGObject
bindings do not do this.

We can avoid problem by reading the audio data directory from the
underlying Gst.Memory object. In this case the Python interpreter is
responsible for copying the data and so it is able to correctly free
the memory after it's no longer needed.

I tested this by calling pyacoustid.fingerprint() on 34 .MP3 files in
sequence, and I saw the following difference:

• memory usage without the patch: 557052 KB
• memory usage with the patch: 52752 KB

The generated acoustid fingerprints were identical with and without the
patch.

[ssssam]

This should fix #62, and possibly #40

[ssssam]

In theory we should be able to use Gst.Buffer.extract() to avoid duplicating memory that we can't free. In practice this leads to a crash. I think that there is a bug in the GStreamer gobject-introspection annotations causing that. I don't have enough information to open a bug against GStreamer right now though.

[sampsyo]

Nice digging! This does indeed seem like it might be the memory leak. And it would, of course, be a bad one—we're leaking all the audio data, not just one metadata record per audio file or something tame like that. I have just a couple of suggestions inline.