DeepSeek-V4.1: common tests, offloading and structure checks
The compressor read its pooling weights straight off its submodules to promote them to float32, so with
offloaded weights it pooled uninitialized memory. The promotion now happens inside the projection itself.
The `shared_states` dict also has to skip device placement, or every layer gets its own copy of it.
`get_image_features` now returns a model output with the vision encoder's hidden states and attentions,
the forwards take the deprecated `past_key_values`, and the tests check the shapes that the compressed KV
and the hyper-connection streams really have.