Fix release race conditions and refactor core flows

[i-surkov]

Test specs added demonstrate race conditions when performing release of values under the same key, when there are multiple concurrent getOrUpdate1, put, and remove operations being performed. Two of these specs (put + remove, and getOrUpdate1 + put + remove) fail on the current master branch.

In short, there were several mix-ups as to which fiber should release which entry and in which case: some entries were being abandoned and not released after being removed, and for some, release was being performed more than once.

[i-surkov]

Changed semantics as discussed with @t3hnar:

1. getOrUpdate1 is responsible for releasing its values that it failed to place in the entryRef. Before, it would expect remove to wait for its deferred and release it. The new rule is more straightforward and allows for simpler reasoning regarding different states between getOrUpdate1, put and remove.
2. remove only performs release of the loaded Values, and ignores Loading ones (returns None.pure[F]). Tests had to be adjusted.
3. put doesn't retry anymore in case of CAS failure when putting new value in entryRef, and simply exit. That helps us to reduce contention, without sacrificing much of consistency.