virtual-sink: Fix a crash when moving the sink to a new master right after setup.

If the virtual sink is moved to a new master right after it has been created,
then the virtual sink input's memblockq can be rewound to a negative read
index. The data written prior to the move starts from index zero, so after the
rewind there's a bit of silence. If the memblockq doesn't have a silence
memchunk set, then pa_memblockq_peek() will return zero in such case, and the
returned memchunk's memblock pointer will be NULL.

That scenario wasn't taken into account in the implementation of
sink_input_pop_cb. Setting a silence memchunk for the memblockq solves this
problem, because pa_memblock_peek() will now return a valid memblock if the
read index happens to point to a hole in the memblockq.

I believe this isn't the best possible solution, though. It doesn't really make
sense to rewind the sink input's memblockq beyond index 0 in the first place,
because now when the stream starts to play to the new master sink, there's some
unnecessary silence before the actual data starts. This is a small problem,
though, and I don't grok the rewinding system well enough to know how to fix
this issue properly.

I went through all files that call pa_memblockq_peek() to see if there are more
similar bugs. play-memblockq.c was the only one that looked to me like it might
be broken in the same way. I didn't try reproducing the bug with
play-memblockq.c, though, so I just added a FIXME comment there.

src/modules/rtp/rtp.h

@@ -40,6 +40,9 @@ typedef struct pa_rtp_context {
 } pa_rtp_context;
 
 pa_rtp_context* pa_rtp_context_init_send(pa_rtp_context *c, int fd, uint32_t ssrc, uint8_t payload, size_t frame_size);
+
+/* If the memblockq doesn't have a silence memchunk set, then the caller must
+ * guarantee that the current read index doesn't point to a hole. */
 int pa_rtp_send(pa_rtp_context *c, size_t size, pa_memblockq *q);
 
 pa_rtp_context* pa_rtp_context_init_recv(pa_rtp_context *c, int fd, size_t frame_size);