introduce this under a new "incubation" sub-package

user to specify as normal DataLoader param

"X, y" is common ordering in ML training usage

allows model training code access query dim sizes

required refactoring to initialize ExperimentDataPipe when needed

cleans up experiment open logic, providing s3 region param

separate the ExperimentDataPipe instantation from the experiment_data_loader()
method, to allow user to specify additional pipe operations, per her use case.

Sparse Tensors are experimental and are currently known not to work with
multiprocessing or LR models. So let's make use of dense Tensors the
default. Use of sparse Tensors now requires the explicit option
`sparse_X=True`.

Also fix bug where dense, non-batched results were returned as 1-row batches,
instead of being unwrapped.

* experimental tests only run with explicit pytest option
* only run experimental tests with Python 3.10 on GHA, since not all experimental package dependencies are available for lower version (e.g. pytorch 2.0.1 not available for Python 3.7)

rename buffer_bytes to soma_buffer_bytes for clarity

- rename ms_name to measurement_name
- remove ExperimentDataPipe num_workers param (unused except for an invalid usage check)

rename layer_name to X_name, to align with Census API get_anndata