…net#4490)

* Ignore hidden columns in AutoML schema checks of validation data

 Found that linux has different order for files being read in than
Windows in order to maintain values, using sort to remove this
discrepency between linux and other OSs

* remove special handling for nix.

* Found special handling of Nix in AutoML as well, now fixed. Also addressed typos.

* use short link to same dataset.

…r. (dotnet#4479)

* Ensure BufferBlocks are completed and empty in RowShufflingTransformer.

If BufferBlock doesn't get completed and drained of its items, it will have non-completed Tasks. When debugging in VS, this will appear to be a memory leak because VS adds all running Tasks to a static Dictionary, and then removes them when the Task is complete. If the Task doesn't get completed, it won't be removed from the Dictionary - thus it looks like a leak.

Note there is no leak when the VS Debugger isn't attached because the non-completed Tasks don't get added to the static Dictionary.

Fix dotnet#4399

* regression trainers export to onnx changes

* Capitalize test

…4462)

* Added onnx export support for several multiclass classifiers

* Separated LightGBM tests for 64-bit and updated baseline for another test

* Fixed bad merge

* Update KeyDataViewType API docs

* Update keyOrdinality description

* Refine addKeyValueAnnotationsAsText definition

* Removed duplicate line

* Revert "Disable tests on macOS that depend on GDI+ library. (dotnet#4492)"

…s from disk (dotnet#4485)

* Changed ComponentCatalog so it would use the most public "initter"
* Changed the visibility of several constructors and create methods so to choose the correct initter when loading models from disk