Reshape the user-facing API around a single MuData input and
mdata.obsm output keys. New entry points:

- MIDAS.setup_mudata(mdata, batch_key=...) registers data setup info
  on the MuData (mdata.uns['_scmidas']) and replaces the previous
  class-level state mutation.
- MIDAS(mdata, configs=..., batch_size=...) constructs the model
  directly from a registered MuData. Instance state instead of
  class-level state — fixes a bug where two MIDAS() instances in one
  process would clobber each other.
- model.get_latent_representation(kind='c'|'u'|'joint') and
  model.get_imputed_values(modality=...) return arrays aligned to
  mdata.obs_names, ready for direct assignment to mdata.obsm /
  mdata.layers.
- model.save(dir) / MIDAS.load(dir, mdata) write/read a model.pt +
  setup.json pair.
- scmidas.datasets.from_dir(dir_path, label_dir=...) loads the
  directory-format datasets (mat/<m>.mtx, mask/<m>.csv,
  feat/feat_dims.toml) into a MuData, including per-batch masks,
  cell labels, and ATAC chromosome chunk dims.
- scmidas.pl.umap(mdata, basis='X_midas', color=[...]) and
  scmidas.pl.modality_grid(model, mdata, label_key=...) wrap the
  common visualizations as one-call helpers and route through a thin
  AnnData wrapper to side-step the current scanpy + MuData plotting
  limitations.
- ATAC binarization is now applied by default whenever 'atac' is
  among the modalities; pass transform={'atac': None} to opt out.

scmidas.integrate() now uses the new path internally and writes the
joint biological latent to mdata.obsm['X_midas'] (configurable via
key_added).

Backwards compatibility:
- MIDAS.configure_data_from_mdata and MIDAS.configure_data_from_dir
  still work — they emit a DeprecationWarning and will be removed
  in 0.4.0.
- save_checkpoint / load_checkpoint still work; new code should use
  save / load.

Other fixes:
- predict(joint_latent=False) no longer raises KeyError: 'z_c'.

15 new pytest cases pin the new API surface (setup, construct,
multi-instance independence, latent/imputation shapes, save/load
round-trip, deprecation warning).