Improve performance of workload generator

[blocked on #50]

## Issue

#50 made the workload generator multithreaded. However, this PR did not fine-tune the generator to make it as effective as possible. There are many different strategies to investigate for improving performance:

* Using non-blocking stubs (see: https://github.com/GoogleCloudPlatform/cloud-spanner-samples/pull/50#discussion_r698914518)
* Using separate channels per worker (see: https://github.com/GoogleCloudPlatform/cloud-spanner-samples/pull/50#discussion_r698914518)
* Changing the `ThreadPoolExecutor` or queue.
* Creating a new handler ([`ThreadPoolExecutor` documentation](https://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ThreadPoolExecutor.html) under rejected tasks) so that failed `executor.submit`s block and retry (see: https://github.com/GoogleCloudPlatform/cloud-spanner-samples/pull/50#discussion_r697725061)
* Removing the queue or adjusting the max queue size (see: https://github.com/GoogleCloudPlatform/cloud-spanner-samples/pull/50#discussion_r698908313)