[Durability] Doc TTL started at sync_write prepare, not at sync_write commit

Build: 6.5.0-3499

Scenario:

1. Create 2 node cluster
2. Default Couchbase bucket with replicas=1
3. Stop Memcached on node-2 (kill -STOP pid)
4. Load 50 docs such that it is inserted into node-1's active Vbucket with doc_ttl=20 and sdk_timeout=60
5. Sleep for 30 seconds
6. Resume Memcached on node-2 (kill -CONT pid)
7. Now all doc_ops succeeds
8. Set flush_param exp_pager_stime = 1 on all nodes
9. Read all the inserted docs

Observation:

8 docs are missing during the read operation, which means the doc's TTL has started (at step #4) before the sync_write commit succeeded (after step #6). This is because we are using 8 threads to insert docs into the server using Java SDK 2.0.0-alpha2.

Note: If the same test case is run with doc_ttl=50, the test passes.

Following doc_keys are missing during the read operation:

[u'test_ttl_docs-51', u'test_ttl_docs-73', u'test_ttl_docs-84', u'test_ttl_docs-30', u'test_ttl_docs-41', u'test_ttl_docs-63', u'test_ttl_docs-6', u'test_ttl_docs-17']

TAF test to reproduce this issue:

/usr/bin/jython  -J-cp jars/sdk3.0/*:jars/* /Users/ashwin/gitRepos/couchbaselabs/TAF/testrunner.py -i node.ini -m rest -tepengine.expiry_maxttl.ExpiryMaxTTL.test_maxttl_with_timeout,nodes_init=2,replicas=1,num_items=50,doc_ttl=30,sdk_timeout=60,durability=MAJORITY