Stardist

- Provide pre-compiled binaries ("wheels") for most platforms, i.e. no need for compilation during installation
- Bug fixes and improvements
- Improve tests, documentation and notebooks
- Internal changes related to testing

Bugfix release, mainly to work around Windows installation issue due to [Windows 10 version 2004 fmod bug](https://developercommunity.visualstudio.com/comments/1243113/view.html).

Also fixes a bug that occurred when training was continued for a model after it was loaded from disk. Thanks Xqua!

This is a big release with lots of (internal) changes:
- Support both TensorFlow 1 and 2 (via [CSBDeep](https://github.com/CSBDeep/CSBDeep) >= 0.6.0)
- Support prediction for very large images via block-wise processing (cf. 36)
- Remove Python 3.5 support and also add Python 3.8 support (when used with TensorFlow 2)
- Add utility function to export 3D results as .obj file (OBJ geometry format)
- Add possibility to train with data that doesn't fit in memory (cf. 57)
- Fix installation bug when numpy is not installed (71)
- Host pre-trained models on GitHub (cf. 48)
- Add new matching metrics
- Other bug fixes and improvements

Changes:
- Support and include pretrained models (cf. 46)
- Improve ImageJ RoI export (subpixel resolution)
- Fix minor bugs related to polygon representations and plotting
- Deprecate Python 3.5. *Starting with StarDist 0.6.0, Python 3.5 will no longer be supported.*

Improve example notebooks (although not part of the python package itself):
- Include evaluation and detection performance after training
- Demonstrate how prediction results can be saved (cf. 45)
- Show and use pretrained models
- Fix some minor bugs in the `other2D` notebooks
- Add data augmentation example for 2D model training
- Notebooks can be run standalone, no need to clone the whole repository anymore

Fix installation bug

This release has many bug fixes and improvements, including:
- Improve non-maximum suppression (faster, interruptible, verbose option)
- Make 2D distance computation more accurate (issue 33)
- Add `export_TF` function to export the model for use in the companion [ImageJ/Fiji plugin](https://imagej.net/StarDist) (cf. #10)
- Fix several bugs that only occurred when using multi-channel input images
- Fix some issues with ImageJ ROI export
- Fix bug to allow using input images without annotated objects (issue 25)
- Add plotting functions to show instance segmentation (2D only)
- Add support for building shared library for C++ code (work in progress)
- Fix potential bug when using constant value label images
- Fix potential bug when using the [`edt`](https://github.com/seung-lab/euclidean-distance-transform-3d) package (deal with non-contiguous arrays)

Minor (but *potentially breaking*) changes:
- Add `train_foreground_only` configuration parameter
- Data type of output label image now `int32` for both 2D and 3D (was `uint16` for 2D before)
- Signature of `augmenter` function changed (now expects a single input/label image pair as input)