Improve executor performance

[ivanpauno]

This PR applies two minor improvements:

• (1) Avoid to repeating some iterations over the same colections.
  Doesn't reduce time complexity, but it reduces the total amount of "operations" done.
• (2) Cache the created wait set and resize it.

The second item seems to be a good improvement as allocation/deallocation is slow.
I have still some tests issues that I'm trying to figure out. Solved!

I have done some profiling, though nothing great (I need to create a demo with more entities to get more interesting results)
To profile:

• Download visualvm, it can be run using ./visualvm_212/bin/visualvm after extracting the archive.
• Run the listener and publisher demos:
  • ros2 run rcljava_examples subscriber_lambda &
  • ros2 run rcljava_examples publisher_lambda &
• Profile the subscriber_lambda
  • Pick the org.ros2.rcljava.examples.subscriber.SubscriberLambda (pid <pid>) entry in Local.
  • Go to the profiler tab and add org.ros2.rcljava.executors.** to the "profile classes:" list.
  • Start running "CPU" profiling.

I have compared between this branch and main:

• In my machine, (1) only improved performance marginally (could've been measurement error).
  It's reasonable because I'm only testing with one subscription available.
• (2) gives a more noticeable improvement.
  waitForWork() went from taking 50% of the time to 37%.
  We still need to test in a more realistic application to confirm.

Note: To make the measurements stable, I have disabled frequently scaling in my machine temporally.
See for example here for details.

[ivanpauno]

I have still some tests issues that I'm trying to figure out.

I knew I was doing something wrong there when I saw the test failures, but I couldn't find the problem 😂.
Fixed it in 80ff53d.

[jacobperron]

LGTM, I'd like to test this out before we merge it.