Fix the collision between padding and the OOV from Categorify

We are facing an issue in ETL part of session-based when we have missing values in string columns. Currently NVTabular pipeline crashes if we do the following:

fill_na_categs = ['category_code', 'brand'] >> nvt.ops.FillMissing(fill_val="unknown")
cat_feats =  fill_na_categs + ['user_id', 'product_id', 'category_id', 'event_type'] >> nvt.ops.Categorify()

We do not have to use FillMissing op for string columns, we can directly use Categorify op, and, then missing values will be mapped to 0 value, and then to null embeddings. But we use 0 value for padding also. Same case happens, when we have OOV each time. This would create a collision issue between padding and the OOV from Categorify op.

No files or tests are named. Reproduce the shown FillMissing and Categorify pipeline, then trace how missing values and OOV categories are mapped relative to sequence padding and null embeddings. Done means missing values and OOV categories no longer collide with the padding value, with regression coverage for both cases.