Remove support start/stop features based on the WITHIN declaration

[johann-petrak]

trafficstars

Turns out this was a rubbish idea and it causes problems when trying to make it work with dense representations where we need to know and control the number of features.

A much better approach is to separate the two things which implicitly happen here:

• Have a PR which creates the indicator features, e.g. "starts-with-TYPE" and "ends-with-TYPE" where we can also control much better how the feature is set
• Then use that feature like any other with an ATTRIBUTE or ATTRIBUTELIST declaration

The WITHIN declaration is still useful to just limit which/how ATTRIBUTELIST features get created, for both dense and sparse representations: for sparse, nothing should get generated outside within, for dense, the default missing values should get assigned.

For ATTRIBUTE, the WITHIN declaration support can be removed as there are easier ways outside of the LearningFramwork to limit the instances to annotations within some other annotation (could also provide a LF PR for that).