fix(csv): CSV update process optimization regression

[ecordell]

I recommend reviewing by commit.

This was an adventure:

• The first commit fixes the actual issue; somewhere in the past we lost the GC optimization that can mark multiple CSVs at once
• When fixing that, I switched to using the typed clients instead of operatorclient's CustomResourceClient
• Because of that, I needed to write a test using the fakes instead of the mocks.
• The test was so much nicer and easier to write than the mocks, that I switched the rest of the tests for the OLM operator to it.
• Because we no longer needed the mocks I was able to delete large chunks of test-only code.
• As part of my refactoring, I broke the re-queuing of CSVs to check for changes. Instead of fixing it short-term, I went ahead and implemented a watch on deployments, which will trigger a sync on any CSVs that own them. This addresses ALM-342
• When verifying my work I fixed some issues in the e2e tests. I'm not 100% sure I removed all flakes, but there may be some overlap with Fix e2e Tests #406

Before merging I want to play with this on a real cluster, but everything looks good locally.

Also:

• e2e tests now run in 2 minutes 32 seconds, down from 7 minutes 20 seconds
• coverage is up ~2% on olm operator

[njhale]

Really cool.