Eagerpy

Fixed PyPI compatibility

New Features

* added `flatten()` with additional `start` and `end` parameters
* added boolean tensor indexing, i.e. `tensor[boolean_mask]`
* other improvements and bug fixes

New Features

* `Tensor.__iter__()` is now implemented (instead of relying on `__getitem__` for iteration, which was inconsistent)
* `norms.l0` now returns the same dtype as the input (previously returned int, being inconsistent with other lp norms)
* `mean` now requires float inputs for all frameworks (previously only for PyTorch and TensorFlow) and throws an appropriate error rather than crashing
* `sum` is now guaranteed to work with int and float tensors as well
* fixed a bug in `sum` for boolean `TensorFlowTensor` instances

New Features

* `Tensor.numpy()` now returns a read-only array for all frameworks (previously this was inconsistent between frameworks and caused bugs because the numpy array may share memory with the actual tensor)

New Features

* reshape now accepts a single `int` (in particular `-1`), not just tuples

New Features

* `Tensor.__getitem__` now supports NumPy arrays (`ndarray`)