docs: explain how the model families diverge, in one place

[Rebreda]

Three families are supported now, and the differences between them were spread across a paragraph in the AMD guide, a docstring, and the code. Someone adding a fourth had no way to know what the complete list of differences was.

New docs/architectures.md states it once.

The table

Every row is a real field on Architecture, so the page maps to the code rather than describing it:

	whisper	moonshine	moonshine_streaming
feature_key	input_features	input_values	input_values
pad_features	False	True	True
pad_to_multiple	–	–	80
supports_language_and_task	True	False	False

The invisible divergences

Two differences are handled by falling back rather than by a field, so they do not appear in the table and would otherwise be undiscoverable:

Offline Moonshine's tokenizer exposes no special tokens at all, so the pad token is read back from the model config. Whisper and streaming Moonshine both have one.

Whisper names its output projection out_proj while both Moonshine variants use o_proj. The default LoRA targets avoid it, so this only matters if you set lora_target_modules yourself.

Also stated plainly, because it caught me out: Lemonade serves moonshine-ai/moonshine-streaming, which is ONNX-only and cannot be fine-tuned. Serving and fine-tuning are different artifacts.

Keeping it honest

Four tests. A stale table here is worse than no table, because it tells someone adding a family that they have covered every difference when they have not. One test enumerates dataclasses.fields(Architecture) and fails if a field is added without a matching row.

Elsewhere

finetune-amd.md's base model section now names the three families and links here, instead of repeating a two-family comparison that was already out of date. README gains the page in its docs table.

386 passed and 14 skipped on 3.11 and on a dev-only install, 406 with the extras. All doc links resolve.