Fix segfault and errors on 2-sided sorted views when data window is invalid

[sc1f]

Retrieving data from a View can be controlled via the start_row/end_row and start_col/end_col arguments to to_json, which are parsed in a similar fashion to Python's range - inclusive of start and stopping at end - 1.

Invalid data windows - where start > end or start == end - are handled by returning empty data. However, in the special case of two-sided sorted views, it did not correctly handle invalid data windows while constructing a column indices vector, and would segfault or throw a ValueError.

This does not raise an issue with the viewer/plugins - the data window requested by perspective-viewer will be valid, even if the viewport is extremely small. However, this interferes with a Websocket testing suite I'm currently building, which randomly generates data windows with the expectation that they are handled correctly by Perspective.

Because there are many situations where a data window can be technically invalid, such as requesting a to_json on a view with 0 rows, or a view with 0 columns (such as the one created when generating autocomplete results for string filtering), I think it's better to have consistent behavior rather than throwing an error.

Changes

• Fixes segfault and ValueError when data windows are invalid for 2-sided sorted views
• Adds tests for all view contexts to check the behavior for invalid data windows are consistent

[texodus]

Looks good!

I think I disagree that this is the desired behavior, since these windows are indisputably "invalid" in the whole-program-correctness, in that negative or zero sized windows are bad API calls from the consumer, and we should inform them as such with a constructive error. However, given the reality of our current perspective-viewer integration, this seems the change-of-least-impact for now.

[texodus]

Thanks for the PR!