RDB Shredder: optimize DAG by excluding count

[chuwy]

Currently we use countAsync to count events in final RDD, but it has proven itself very inefficient in regards to cache and also not being async. We can replace countAsync with event counter implemented as Spark accumulator to get rid of additional job.

Usually counters implemented on top of accumulators is a bad idea because they can be executed more than once, but we decided it's a reasonable trade-off because:

1. We never need an exact number of events - it's purely informational
2. The chance of another attempt is very small at the point where we increment

This commit aslo simplifies the DAG by joining good and bad data output into a single step. And last, it saves shredded type only if it has been successfully shredded.