Latent codec: more natural initialization

[YodaEmbedding]

• docs: updates for sphinx-book-theme==1.0.0 (fixes warnings)
• latent codec: Use self-documenting keyword arguments instead of _setdefault et al.
• latent codec: Got rid of redundant N argument.

To do:

• Test everything again.

[fracape]

entropy_bottleneck=EntropyBottleneck(M) ?

I see, for cheng / sensetime, it's still N

[fracape]

entropy_bottleneck=EntropyBottleneck(M) ?

[fracape]

What is the rationale to init an entropybottleneck with 0 channels, can you extend it later?

In gainhyperlatentcodec, you init with
self.entropy_bottleneck = entropy_bottleneck or EntropyBottleneck()

[YodaEmbedding]

You're right; this was a bit silly. I wanted it to type-check. Possible alternatives:

1. Force callers to provide entropy_bottleneck:

   def __init__(
       self,
       entropy_bottleneck: Optional[EntropyBottleneck] = None,
   ):
       assert entropy_bottleneck is not None
       self.entropy_bottleneck = entropy_bottleneck

   Usage:

   HyperLatentCodec(
       entropy_bottleneck=EntropyBottleneck(N),
       h_a=h_a,
       h_s=h_s,
   )

2. Restore the z_channels argument:

   def __init__(
       self,
       z_channels: Optional[int] = None,
       entropy_bottleneck: Optional[EntropyBottleneck] = None,
   ):
       self.entropy_bottleneck = entropy_bottleneck or EntropyBottleneck(z_channels)  # type: ignore

   Usage:

   HyperLatentCodec(
       z_channels=N,
       h_a=h_a,
       h_s=h_s,
   )

I think (1) seems simplest / most explicit / consistent with users already providing h_a=h_a, h_s=h_s.