Fix handling of users with multiple identities

[tgoyne]

ROS users were identified by the pair (identity, auth_url), and in the v10 port this was translated to (user_id, provider_type), but that isn't actually how Atlas users work. An App is the equivalent of the old auth_url, and within an App user_id uniquely identifies a user. Users don't actually have a "provider type" at all: users have one or more identities, each of which has a provider type, and the same SyncUser should be used regardless of which identity was used to log in.

This had surprisingly mild consequences, but was visibly broken in a few ways. Logging into the same user using two different identities resulted in two SyncUsers with the same user id. Opening the same partition (or any flx Realm) with both users resulted in both using whichever SyncUser happened to open it first, and the second would not create a SyncSession. This meant that most of the potential bad things (such as creating two session for one file) didn't actually happen, but the second user's list of sessions would be "incorrect", and removing one of the users would remove the incorrect set of local files.

Linking identities to anonymous users resulted in the user still being treated as anonymous. The primary negative effect of this was that linking an identity, logging out, then logging back in would result in the user being removed entirely in between, forcing the re-download of all Realms (and the loss of any un-uploaded data).

This bumps the schema version of the metadata Realm primarily for the sake of recovery on metadata Realms in invalid states: the migration block handles the case of multiple users with the same user id and unifies them. Since this needs a migration anyway, it fixes some incidental problems with the metadata Realm's schema which weren't fixable without a migration: the marked_for_deletion column was redundant with the Removed state, identity objects were leaked due to not being embedded, and the local_uuid field is no longer used for anything other than opening files written by early v10 versions, so there's no need to populate it for new users.

Initializing a sync manager opened the metadata Realm, read the persisted users, and then wrote the persisted users back to the Realm using three separate write transactions per user. It now just reads the data and doesn't write anything.

SyncUser's internal API permitted doing some invalid state transitions such as marking a user as logged in despite it not having any tokens. The two specific things which definitely could break was logging out a user while the profile was being updated and logging out a user while the access token was being refreshed, but there may have been more things. There also just was some dead code trying to support state transitions which could never happen.

[tgoyne]

This unfortunately grew to be pretty large in the process of making tsan happy, but hopefully the commits make enough sense individually that it's not too awful to review. There's no particular hurry here; this is just some incidental bug fixes extracted out of my multiprocess sync work.

[coveralls-official]

Pull Request Test Coverage Report for Build github_pull_request_275077

• 799 of 830 (96.27%) changed or added relevant lines in 24 files are covered.
• 31 unchanged lines in 13 files lost coverage.
• Overall coverage increased (+0.07%) to 91.237%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/realm/object-store/sync/impl/sync_metadata.cpp	199	200	99.5%
test/object-store/sync/app.cpp	82	83	98.8%
test/object-store/util/sync/sync_test_utils.cpp	4	5	80.0%
src/realm/object-store/sync/sync_user.cpp	72	78	92.31%
test/object-store/sync/metadata.cpp	90	96	93.75%
test/object-store/util/unit_test_transport.cpp	128	144	88.89%

Files with Coverage Reduction	New Missed Lines	%
src/realm/object-store/c_api/app.cpp	1	28.81%
src/realm/object-store/sync/app.cpp	1	93.3%
test/object-store/sync/session/session.cpp	1	98.5%
test/test_dictionary.cpp	1	99.85%
src/realm/object-store/shared_realm.cpp	2	94.59%
src/realm/object-store/sync/sync_user.cpp	2	90.57%
src/realm/sync/instruction_applier.cpp	2	70.35%
src/realm/sync/noinst/client_impl_base.cpp	2	85.54%
src/realm/sort_descriptor.cpp	3	94.33%
src/realm/sync/network/network.cpp	3	89.87%

Totals
Change from base Build 1680:	0.07%
Covered Lines:	233774
Relevant Lines:	256227

---

💛 - Coveralls

[ironage]

should this test wait for the access token response before checking the user state?

[ironage]

oh never mind, we use a synchronous transport

[ironage]

This is a much needed clean up of a neglected area of the code. 👍
Adding some defensive checks (or assertions) in log_out_user would be bonus, but not blocking.