Only run raster operations after image sources have loaded

[tschaub]

The raster tests were not properly cleaning up after themselves. Listeners for afteroperations were getting called a second time after we had called done() in the tests. This made it look like the wrong tests were failing.

This cleans up the raster tests by calling obj.dispose() on reused test objects and only listening for events once where we call done() in the listener.

There may still be an issue introduced in 98b823c with operations being run more often than expected. But we should create a specific test that demonstrates the issue if one can be found. The existing test failures were the result of sloppy test running.

update - this should now fix #4893.

[tschaub]

In this branch, tests fail before 24b7b27 and pass after it. So there was really only one change needed. But the other changes made the problem clearer.

[ahocevar]

This looks much clearer now. Thanks @tschaub, and please merge when Travis is happy.

[tschaub]

Well, Chrome is still not happy. I lied about 24b7b27 making all tests pass. The first setOperation() test is still failing. This looks legit and could be the root of the problem.

[ahocevar]

@tschaub, 9983586 fixes the issue. I haven't tested yet whether this has unwanted side effects, but the problem is that previously the renderer's getImage() method was called, which returned null if the source was not ready. Now, composeFrame() is called, and we do that without checking the return value from prepareFrame() to see whether there is something to render at all.

[tschaub]

Thanks for the underlying fix @ahocevar. To summarize for others:

• prior to 98b823c, the raster source would not run pixel operations before any data had been rendered (see https://github.com/openlayers/ol3/blob/557372203f25c1502a16d52b7e1000eb7e5b5265/src/ol/source/rastersource.js#L347-L355).
• the changes in 98b823c improved things by not rendering to an intermediate canvas, but there was no way for image layers to indicate that no work needed to be done before an image had loaded
• the failing tests noted in ol.source.Raster tests are failing #4893 were misleading because the tests were not cleaning up after themselves, but the underlying issue is that now raster operations were run once before an image loaded and once after
• the first few commits here cleaned up the tests and made it clearer which one was a legitimate failure
• f2d5d57 makes it so the renderer knows there is no more work to be done if an image is not loaded (fixing the real issue)

So, it is nice that we have a test that actually covers the problem. But it is a bummer that these tests are not run as part of our CI setup. Travis/Sauce integration can help - though it is a bit involved to get it working for 3rd party pull requests.