improve __call__ performance benchmarking

[daniel-sanche]

In #527, we improved the performance of _GapicCallable's call method, and added a simple benchmark unit test to prevent performance from regressing too much

The test was very simple, using timeit to measure the time to run 10,000 calls, and dailing if it was slower than expected. This doesn't take performance of the worker into account, so the test may flake in different environments. To prevent this, we left a large buffer, but this means smaller regressions won't be caught by the test

@parthea suggested opening this issue to investigate improving the performance tests in the future. If we could test the performance of the old revision and the suggested change as part of the same test, we could make much better assertions on the performance of the library

CC @ohmayr

[daniel-sanche]

Related: #542 (comment)

[vchudnov-g]

See also my suggestion for capturing tighter limits per platform: have a dict of platform to expected performance threshold, and have the failing assertion print out both the time and the current plaatform so we can add that pair to the table for future runs.