Pietoolbelt

* Multilabel stratification
* Improved MLFlow monitor

New features:
* Add predefined augmentations for detection and segmentation steps
* Start implementing data flow control interface
* Add visualizer for detection task

New features:
* Added `steps.segmentation.Inference`
* Added image TTA operations and support of it in inference
* Added `models.utils.ModelsContainer`. Now we can push set of models into it, specify the blending method and use it as a single model

Breaking changes
* The `datasets.DatasetStratification` was moved to `steps.stratification`

Fixes
* Dataset stratification on low data
* Colormat visualizer now doesn't change the color of unsegmented image part
* `MasksComposer` now return mask with values equals 0 or 1 for every class after masks compose

Functional
* Datasets
* `datasets.stratification` - stratification by histogram
* `datasets.utils` - set of datasets constructors that
* Losses
* `losses.common` - losses utils
* `losses.regression` - regression losses
* `losses.segmentation` - losses for single and multi-class segmentation
* `losses.detection` - losses for detection task
* Metrics
* `metrics.common` - common utils for metrics
* CPU - metrics, that calculates by `numpy`
* `metrics.cpu.classification` - classification metrics
* `metrics.cpu.detection` - detection metrics
* `metrics.cpu.regression` - regression metrics
* `metrics.cpu.segmentation` - segmentation metrics
* Torch - metrics, that calculates by `torch`
* `metrics.torch.classification` - classification metrics
* `metrics.torch.detection` - detection metrics
* `metrics.torch.regression` - regression metrics
* `metrics.torch.segmentation` - segmentation metrics
* Models
* `decoders.unet` - UNet decoder, that automatically constructs by encoder
* `encoders.common` - basic interfaces for encoders
* `encoders.inception` - Inceptionv3 encoder
* `encoders.mobile_net` - MobileNetv2 encoder
* `encoders.resnet` - ResNet encoders
* `albunet` - albunet model
* `utils` - models utils
* `weights_storage` - pretrained weights storage
* Pipeline steps
* `regression.train` - train step for regression task
* `regression.bagging` - bagging step for regression task
* `img_matcher` - image comparision and matching tool based on descriptors
* `mask_composer` - mask composer tools that can effectively combine masks for regular, instance or multiclass segmentation
* `utils` - some utils
* `viz` - image visualisation tools