Refactor gen resources

[scotts]

Two refactors:

1. Resource generation, generate_reference_resources.py, is now more deeply integrated with our testing utils. This allows us to greatly simplify the resource generation, and it keeps the actual definitions in utils.py. That is, the video we define in utils.py defines the filename, and it has the logic for how to name reference frames.
2. Moves around definitions in utils.py so that we group all video and audio test files together.

[NicolasHug]

LGTM, the FFmpeg4 issue is known

Dumb Q, but why do we generate those test resources across all FFmpeg versions in our CI?

https://github.com/meta-pytorch/torchcodec/blame/1befed7063e62b803da0927567d34d18c62a54a3/.github/workflows/reference_resources.yaml#L56

IIRC, we've only been using FFmpeg 4 for the reference generation so far?

[scotts]

@NicolasHug, good question. While we generated all of these resources originally with FFmpeg 4, they're actually stable across all versions we test against. For example, I've been using FFmpeg 6 for all of my testing the past several months, and when I re-run the resource generation script, everything comes out clean - we're bit-for-bit compatible with what we originally checked in.

This may seem surprising at first, but if we expect TorchCodec to be bit-for-bit compatible across versions, it's not surprising if the ffmpeg CLI is as well. I think that at the least, we'd want to know if our resource generation script was giving different results across FFMpeg versions.

I think we do have a comment somewhere about using FFmpeg 4. We should probably delete it, if we haven't already.