Kwimage

Changed
* Torch and pandas are now optional

Added
* `kwimage.draw_text_on_image` now support `halign`
* `kwimage.Boxes.draw_on` now supports different colors for each box.

Changed
* Removed explicit Python 3.5 support. Note, 3.5 should still still work using
the universal `py2.py3-none-any` wheel.

Fixed
* Issue with `Detections.from_coco` with keypoint categories
* Fixed `kwimage.Boxes.draw_on` when images are non-contiguous

Added
* Add `to_boxes` to MultiPolygon

Changed
* More methods in the Mask object should work without the c-extensions being built.
* The `Mask.to_coco` method now returns a format based on the native encoding.
* Support for the new is preferred "ltrb" format over "tlbr" which will eventually be deprecated.
* No longer publishing wheels / CI testing for Python 2.7

Fixed
* Fix bug with check for turbojpeg in imread.
* `subpixel_slice` now works with non-integer slices

Added
* `Detections.from_coco_annots` now accepts `dset` keyword argument that
attempts to fill the category_id coco field.
* `Boxes.iooas` - intersection-over-other-area
* `kwimage.imread` - now has a turbojpeg backend.

Fixed
* Fix bug in `Detections.from_coco_annots` when categories are null.
* Fix bug `Detections.compress` when flags are in tensor space.

Changed
* `kwimage.imwrite` now always returns the file path that was written to
regardless of the backend.

Added
* `draw_line_segments_on_image`
* Boxes.scale now accepts `about` keyword arg (can use to scale about center).
* Boxes.warp now accepts matrices and does inexact corner warping
* kwimage structures `warp` function now accepts a generic callable for mapping array-based points.
* add ``normalize`` function for intensity rebalance.

Changed
* Renamed `_rectify_interpolation` to `_coerce_interpolation`. Old function is deprecated and removed in the future.
* `_coerce_interpolation` now accepts strings for fallback interpolation flags.
* `Detections.from_coco_annots` now returns classes as ndsampler.CategoryTree when possible

Added
* Added `im_filter` module with `fourier_mask` function.
* Add "amazon" demo image