Merge pull request theislab#89 from theislab/tf2_compatibility

- structural changes to batchglm necessary before tf2 code merge
- introduction of numpy backend

fixed constraint unit test

fixed bug that division was forced as numpy even though x could be cs…

…r_matrix

using np.divide instead of / made the code very slow (seemingly arresting) here.

removed remaining xarrax references

changed default estimator batch size to 512

relaxed numpy and tf requirements

made access to error codes and niter of estimator nicer

adapted to anndata not exporting anndata.base anymore

Adapting according to discussion at scverse/anndata#174.

adapted to anndata not exporting anndata.base anymore

Adapting according to discussion at scverse/anndata#174.

added dataArray support to constraints building