Remove sparse tensor output type for list features

[oliverholworthy]

Remove sparse tensor output type for list features

Motivation

The value count attributes of columns in dataset.schema currently controls the output type of list columns.

With the addition of shape in the schema NVIDIA-Merlin/Merlin#813 we're going to start seeing value counts specified more. This will result in unexpected output type's if previously a value count was not specified.

We also currently have a possibility of output type sparse tensor which doesn't have a clear use-case and appears to be an implementation detail of padding ragged columns to dense.

Current

• Dense Tensor
  • value_count.max specified and is_ragged=False
• Sparse Tensor
  • value_count.max specified and is_ragged=True
• Ragged Representation (values, offsets)
  • value_count.max not specified and is_ragged=True

After - With this PR

• Ragged Representation
  Always returns this for all list columns. value count does not influence output type.

Planning a follow-up / independent change in #97 to return a dense tensor if the schema specifies that a column is of fixed size (is_ragged=False).