Fix tfidf using the wrong layer

[rcannood]

TF-IDF seems to be using adata.X for some of its computations even when from_layer is defined.

This leads to the following error when .X is not defined.

Input:

AnnData object with n_obs × n_vars = 600 × 1500
    obs: 'tech', 'celltype', 'size_factors', 'n_counts', 'cell_type', 'batch'
    var: 'n_cells', 'feature_name'
    uns: '_from_cache', 'data_reference', 'data_url', 'dataset_description', 'dataset_id', 'dataset_name', 'dataset_organism', 'dataset_reference', 'dataset_summary', 'dataset_url', 'var_names_all'
    layers: 'counts'

Code:

normalized_counts = ac.pp.tfidf(
  adata,
  from_layer="counts",
  to_layer="tfidf"
)

Error:

Traceback (most recent call last):
  File "/tmp/viash-run-atac_tfidf-6y8F0o.py", line 39, in <module>
    normalized_counts = ac.pp.tfidf(
                        ^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/muon/_atac/preproc.py", line 106, in tfidf
    idf = np.asarray(adata.shape[0] / adata.X.sum(axis=0)).reshape(-1)
                                      ^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'sum'

Not only that, but it could mean that muon is accidentally computing incorrect results when .X and from_layer is used.