DeepSeek-V4.1: resolve the model structure findings
The four attention classes are now one class that branches on the layer
type, like DeepSeek-V4 does, so there is no subclass chain and no class
dict any more.
The model owns the only rotary embedding. It computes the RoPE of the
compressed entries for each compress ratio and passes it down, instead of
the compressing attention layers building a rotary module each.
`set_engram_token_map` is a single method on the PreTrainedModel instead
of one per model class. It still needs a linter ignore: it fills a saved
buffer that the checkpoints leave out, not a hyperparameter.
Also merges main, which dropped `keys_to_ignore_at_inference` from the
parent config.