Improving column purpose detection for sparse datasets

AutoML does poorly on a few text datasets. For example, a text dataset we benchmark on has an accuracy of 0.60 vs. an expected accuracy of 0.85.

This is caused by us detecting the text columns columns as **categorical** instead of **free text**. For the this dataset, this is due to the text column being 84% blank (a sparsely filled out column).

**To fix:**
We need to detect the column purpose only on the set (non-blank) values.

Recommend subtracting the blank values from `data.Count`:
https://github.com/dotnet/machinelearning/blob/227da9d7db2ce80b073cc64bfd067b04e6189de1/src/Microsoft.ML.AutoML/ColumnInference/PurposeInference.cs#L148-L158

Currently `avgLength`, `cardinalityRatio`, `avgSpaces` are artificially lower due to the missing values.