Remove secondary mappings for flattening operators

For operators like mergeMap, switchMap, concatMap, etc, we have a secondary selector. It's not used a lot in practice that I've seen. Originally, it added a lot of value in terms of perf because it prevented a closure... however, in practice it makes it harder to isolate errors from observation chains, and in a world with Turbofan, I'm not sure the performance benefit is there. We could eliminate a good amount of code in the library written to support this.

Use and Alternatives

Observable.of(1, 2, 3).mergeMap(a => range(0, a), (a, b) => a + b);

/// which is mostly the same as

Observable.of(1, 2, 3).mergeMap(a => range(0, a).map(b => a + b))

// however if there's an error in that second map, you're forced to kill the whole observable...

Observable.of(1, 2, 3)
  .mergeMap(a => range(0, a), (a, b) => { throw new Error('lol') })
  .catch(err => empty())

// rather than this, which is much more common

Observable.of(1, 2, 3)
  .mergeMap(
    a => range(0, a)
      .map(b => a + b)
      .catch(err => empty())
  )

Proposed Change

Eliminate the second argument to mergeMap, concatMap, switchMap, et. al. and instruct people to use .map within the projection function if they want a secondary level of mapping.

Start by locating the flattening operators named in the issue, including mergeMap, concatMap, and switchMap, and inspect how their secondary selector arguments are exposed and tested. The work is done when those secondary arguments are eliminated consistently and the documented alternative is to use map within the projection function.