Avoid document body unmarshalling on changes requests

[adamcfraser]

Currently changes requests that specify style=all_docs have unnecessary performance overhead when retrieving branched documents.

changes.addDocToChangeEntry will retrieve the document from the bucket (to identify the set of leaf revisions). In the process it will unmarshal the entire doc.

Two enhancements:

• Store the list of leaf rev ids in the revision cache along with the active revision, to avoid retrieving the document from the bucket when it's resident in the rev cache
• When bucket retrieval is required, unmarshal only the rev history, not the full doc body.

[adamcfraser]

For the second item ("When bucket retrieval is required, unmarshal only the rev history, not the full doc body.") - in the context of this ticket, this is specifically about the call to db.GetDoc in changes.addDocToChangeEntry.

One option would be to add a new db method (like 'GetDocSyncData') that only returns the sync data.

Another would be to modify the document struct to always lazy unmarshal the document body. This has the potential for more benefits across the entire codebase, but would require more significant refactoring - it might make more sense to hold that refactoring until 2.0, when the sync data moves to xattrs.

[tleyden]

Notes to self:

1. Get doc bytes by calling GetRaw
2. Call UnmarshalDocumentSyncData

Benchmark:

1. Pre-timer setup: Create 2k docs of size 50k, 1000 keys with branches, 1 parent + 2 child branches -- doesn't matter which API .. bucket api
2. Start the timer
3. Changes request of all docs (could also do GetDoc call, but misses other possible things). One shot, .. etc

[tleyden]

Test results

Fix Status	Millisconds/Op	Bytes/Op	Allocations/Op
Without Fix	5074	699042928	1031780
Without Fix	5127	699030160	1031734
Without Fix	4689	699035600	1031753
With Fix	1366	341696136	493655
With Fix	1334	341695456	493651
With Fix	1307	341695536	493652

Without fix: 8ef9a00
With fix: 341bf80

So it looks like a ~3.5X speedup