Ndsampler

Added
* The target dictionary can now contain `space_slice`, `time_slice` or `gids`

Added
* The target dictionary in `CocoSampler.load_sample` can now accept a `vidid`
and will return a 3D sample.

Changed
* New version will kwarray >= 0.5.16
* The CocoSampler now defaults to backend=None instead of backend="auto".
* The `load_sample` signature was modified and deprecated args were removed.

Added
* The target dictionary in `CocoSampler.load_sample` can now accept a `channels`
to grab specified channels in multispectral images.


Removed

* `_hack_old_names` from `AbstractFrames`
* `CocoDataset` which was moved to `kwcoco`
* `toypatterns` which was moved to `kwcoco`
* `stats_dict`, which now exists in `kwarray`
* `ndsampler.util` module, use `ndsampler.utils.util_misc` instead.
* `ndsampler.util_futures`, which now lives in `ndsampler.utils.util_futures`
* `make_demo_coco`, use `kwcoco toydata` instead.

* `Frames._lookup_hashid`
* `Frames._lookup_gpath`
* `Frames._gnames`

Modified

* Overhaul of `AbstractFrames` for better auxiliary data integration

Changed
* Removed Python 3.5 support
* No longer using protocol 2 in the Cacher
* Better `LazyGDalFrameFile.demo` classmethod.

Fixed
* Bug in accessing the LRU

Fixed
* Minor compatibility fixes for `ndsampler.CategoryTree` and `kwcoco.CategoryTree`

Added
* Can now call `CocoSampler.load_item` with target dictionary only containing
an annotation id.

Changed
* CategoryTree now uses a mixin class to separate class storage from hierarchy
computations.

Fixed
* In rare instances a floating point error would cause `load_sample` to raise
an `AssertionError`. This is now checked for and corrected.