[ML] Consider supporting sub-second bucket_spans for sensor/IoT data use cases

Please reference https://discuss.elastic.co/t/machine-learning-bucket-span-millisecond-precision/197908

User wants to analyze data at 1ms granularity.

Unfortunately, no meaningful description of the use case and applicability to real-time operation. It really seems like an offline data analysis that can be solved with a time-scaling workaround that is described in the thread.