Audio decoding TODOs

[NicolasHug]

Basic support was added in #538, we need the following before we make this public:

• enable backwards seeks Allow audio decoder to seek backwards #550
• enable audio formats other than fltp Support all audio formats by converting to FLTP #556
  • maybe we don't always need to convert to fltp, especially flt (non-planar) formats, we can probably just output tensors with a different layout or so something smarter that avoids a copy.
• enable a user-defined sample_rate Allow sample_rate parameter to audio decoder #551
• expose a public method in AudioDecoder Add AudioDecoder.get_samples_played_in_range() public method #555
• maybe, maybe not: something like the normalize parameter of the torchaudio reader, which allows users to specify whether they want a float tensor in [-1, 1], or a tensor with the same dtype as the audio format. We'll figure that out later. We'll probably always return float tensors by default anyway.
• perf: try to pre-allocate the output tensor and save copies