fix: fix flaky tests and non blocking semaphore

[mutianf]

Fix for #1359

In NonBlockingSemaphore, updating both limit and avaliablePermits in increase and decrease limits may cause problems. If the limit is lowered but avaliablePermits is not updated yet, and release is called, this statement in release: if (availablePermits.compareAndSet(old, Math.min(old + permits, limit.get()))) may update the availablePermits incorrectly.

Since we're already tracking the limit in the semaphore class, instead of tracking what's available, we could track how many permits are acquired. So When we're adjusting the limit, we won't need to update both variables and should avoid this problem.

After the change the test succeeded in all 100 runs:

//gax:com.google.api.gax.batching.FlowControllerTest                     PASSED in 1.7s
  Stats over 100 runs: max = 1.7s, min = 0.9s, avg = 1.2s, dev = 0.1s

[vam-google]

LGTM, but can you please confirm that it is still green on 1,000 and maybe 10,000 runs?

[mutianf]

10000 runs with flag

--test_filter=com.google.api.gax.batching.FlowControllerTest#
--runs_per_test=10000
--cache_test_results=no
--sandbox_tmpfs_path=/tmp

all passed

INFO: 10001 processes: 1 internal, 10000 linux-sandbox.
INFO: Build completed successfully, 10001 total actions
//gax:com.google.api.gax.batching.FlowControllerTest                     PASSED in 2.4s
  Stats over 10000 runs: max = 2.4s, min = 0.9s, avg = 1.3s, dev = 0.2s