perf: improve bson iteration performance

[tychoish]

Minor: this avoids casting bson into values until we need to. This
means that if a document in a stream has a field which is not in the
schema, we just skip it.

[vrongmeal]

How does this affect the performance. I can't figure out where do we "skip" if the field doesn't exist?

[tychoish]

How does this affect the performance. I can't figure out where do we "skip" if the field doesn't exist?

It depends on the data, but it makes the process of iterating through a document lazier. If you have a schema that was (a int, b text) but a document that was {"a":1,"c":<...>,"b":"foo"} you could skip reading the c field entirely. Because all fields are length prefixed you just skip ahead... in the previous iterator it would need to parse the value of the field even if we weren't going to use it.