The spec is a bit unclear whether this is even valid for things like images, however the parameter doesn't really make sense for non-text data anyways.

Synapse currently (for better or worse) treats the presence as a charset as a completely different media type and thus refuses to thumbnail it. 

assume no orientation on bad exif information

Request remote media with r0 instead of v1

Fixes https://github.com/turt2live/matrix-media-repo/issues/267

Fixes https://github.com/turt2live/matrix-media-repo/issues/272

Fixes https://github.com/turt2live/matrix-media-repo/issues/271 at the same time

We make our own ReadCloser+ReadSeeker implementation because a NopCloser actually destroys the seeker implementation.

This needs to be expanded to cover more of what ffmpeg covers. This does not support animation for a few reasons:
* Most clients which will be using a video thumbnail will be using it in such a way where even a GIF isn't supported.
* Clients rely on the thumbnail endpoint returning images, which can lead to resource inefficiencies if we return GIFs.
* Returning more optimized thumbnails is not feasible for compatibility with most clients.

This is a step towards video processing, maybe.

Instead of buffering all the export files into memory (which could be hundreds of gb) we buffer them in one at a time and hope that they don't contain extra information.

On a 171mb export, this change used 45mb of memory at most. Previously the task would have required about 1gb. For larger (200gb+ exports) this is obviously better.