Rework handling of skipped features in replay

[elemoine]

There is a bug in the handling of skipped features in the canvas replay when replaying the hit detection instructions. This PR aims at fixing the bug.

Description of the issue:

Currently the canvas replay uses the instructionIndices_ object to store instruction index/feature uid pairs and be able to get the feature uid for an (begin geometry) instruction and know when a feature/geometry should be skipped. See https://github.com/openlayers/ol3/blob/master/src/ol/render/canvas/canvasreplay.js#L256-L261.

But this doesn't work when replaying the batch for feature detection. For feature detection the replay indeed uses a specific instruction arrays (hitDetectionInstructions), so the instruction indices stored in instructionIndices_ are not valid for hit detection. So with the current code, when replaying the instructions for hit detection, features that should not be skipped may be skipped, and features that should be skipped may not be skipped.

The problem can easily be reproduced with just three point features in a vector layer and a select interaction. It's much more difficult to reproduce when the vector layer has many features (because you must be unlucky enough to find a problematic pair of features).

This PR fixes the bug by removing the instructionIndices_ object entirely and storing the references to features in the begin geometry instructions. This simplifies the code and makes "skip feature" work with hit detection as well.

[tschaub]

A description of the bug would be nice (when you get a chance).

[elemoine]

Yes, I just wanted to quickly push my work and open this PR. I will write a proper description of the issue and how my patch addresses it.

[elemoine]

Description updated. Ready for review.

[elemoine]

Please review. This is a major bug that we need fixed in master.

[tschaub]

This looks like a nice fix to me. Please merge.

[elemoine]

Thanks for the review.