The reason is that the DTS decoder has to decode at least one frame
to set correct channel_mask. For raw DTS, ffmpeg decode one frame
to get stream parameters. For mkv, ffmpeg already have all parameters
(except channel_mask), and thus, it don't try to decode a frame
before displaying stream parameters.
The follwing patch add channel_layout to the list of essential
parameters. This forces ffmpeg to decode an audio frame to get
channel_layout before using it. It fixes the issue with above mentionned
sample file. But I guess it could have bad effects with other codecs
which still don't implement channel_layout (untested).