CUDA support in blocks - #1599

Everything downstream of an AVFrame's owner took `UniqueAVFrame&` or
`const UniqueAVFrame&`. That's a constraint on the caller's storage rather
than a statement about what the function does, and it has two costs.

A non-const `UniqueAVFrame&` lets a callee take ownership of the caller's
frame. Both CUDA interfaces did: BetaCudaDeviceInterface moved out of it,
CudaDeviceInterface reassigned it. That is invisible under
SingleStreamDecoder, whose frame is a loop local that dies right after
conversion, but the building-block ops own their frame in a handle that
outlives the call, so the frame gets freed twice.

`const UniqueAVFrame&` doesn't allow the steal, but it still forces anyone
holding a plain AVFrame* -- which is what the ops' tensor handle really is
-- to manufacture a unique_ptr just to make the call, and manufacturing a
second owner for an already-owned object is its own bug factory.

So: functions that only look at a frame now take `const AVFrame&`, and the
one that writes to it takes `AVFrame&`. Ownership stays with whoever
actually owns the frame. Producers (receive_frame) keep `UniqueAVFrame&`
because they really do hand back ownership.

With that, the ops layer needs no ownership sleight-of-hand:
wrap_pointer_to_tensor() gains a deleter parameter, so the one generic
handle covers Demuxer/PacketDecoder/ColorConverter and the FFmpeg types
too, and both bespoke wrap_*_pointer_to_tensor() functions go away.
Demuxer::next_packet() returns UniqueAVPacket rather than a raw pointer
plus a comment telling the caller to free it.

The encoder is left alone: it owns and mutates its frames, and it uses a
null frame as the flush signal, so a reference is the wrong shape there.

Two things only show up when building against the older FFmpeg headers.

get_num_channels() was patching av_frame.channel_layout when FFmpeg 4 left
it unset, so it was a mutator wearing an observer's name, and the layout
fix-up was a side effect that swresample setup silently relied on. Pull the
fix-up into get_channel_layout() and call that from the two places that
actually need a layout; get_num_channels() just counts.

swr_alloc_set_opts2() only became const-correct in FFmpeg 6 (libswresample
4.12). Cast for the older headers, which don't modify the layout either.

…_cuda_mine

…color_conversion