Support for multiple matrices and improving construction of TileDB objects

[jkanche]

This PR introduces major improvements to matrix handling, storage, and performance, including support for multiple matrices in H5AD/AnnData workflows and optimizations for ingestion and querying.

Support for multiple matrices:

• Both build_cellarrdataset and CellArrDataset now support multiple matrices. During ingestion, a TileDB group called "assays" is created to store all matrices, along with group-level metadata.

This may introduce breaking changes with the default parameters based on how these classes are used. Previously to build the TileDB files:

dataset = build_cellarrdataset(
    output_path=tempdir,
    files=[adata1, adata2],
    matrix_options=MatrixOptions(matrix_name="counts", dtype=np.int16),
    num_threads=2,
)

Now you may provide a list of matrix options for each layers in the files.

dataset = build_cellarrdataset(
    output_path=tempdir,
    files=[adata1, adata2],
    matrix_options=[
        MatrixOptions(matrix_name="counts", dtype=np.int16),
        MatrixOptions(matrix_name="log-norm", dtype=np.float32),
    ],
    num_threads=2,
)

Querying follows a similar structure:

cd = CellArrDataset(
    dataset_path=tempdir,
    assay_tiledb_group="assays",
    assay_uri=["counts", "log-norm"]
)

assay_uri is relative to assay_tiledb_group. For backwards compatibility, assay_tiledb_group can be an empty string.

• Parallelized ingestion:
  The build process now uses num_threads to ingest matrices concurrently. Two new columns in the sample metadata, cellarr_sample_start_index and cellarr_sample_end_index, track sample offsets, improving matrix processing.

  • Note: The process pool uses the spawn method on UNIX systems, which may affect usage on windows machines.

• TileDB query condition fixes:
  Fixed a few issues with fill values represented as bytes (seems to be common when ascii is used as the column type) and in general filtering operations on TileDB Dataframes.

• Index remapping:
  Improved remapping of indices from sliced TileDB arrays for both dense and sparse matrices. This is not a user facing function but an internal slicing operation.

• Get a sample:
  Added a method to access all cells for a particular sample. you can either provide an index or a sample id.

sample_1_slice = cd.get_cells_for_sample(0)

• Other updates to documentation, tutorials, the README, and additional tests.

[jkanche]

@tony-kuo and @keviny2 i'll update and merge this tomorrow, but good to get another set of eyes here

[jkanche]

This PR has dependencies in #61 and #60. Will merge this PR first to keep it small and avoid merge conflicts later.