Fix ingester startup with a normalised ring

[bboreham]

Fixes #1311

[csmarchbanks]

Code looks good, but would it be possible to add a test case in/similar to TestRingNormaliseMigration?

[bboreham]

I added a test which led me to realise my original approach was flawed.
New approach is to do the denormalise every time we need TokensFor() an ingester, which is in initRing() and autoJoin().
(and hence I realise autoJoin() didn't check for clashing tokens in normalised mode)

migrateRing() is now rather poorly named, but I felt we could move that later.