autofill: Understand how tombstones in the mirror are being used in practice

The fix for #5020 was speculative, as I could not understand why invalid records in the mirror could have such a large impact - it should only have been hit by people to disconnected and reconnected sync or some other edge-cases. However, @lougeniaC64 reports that telemetry shows it did fix the problem - which makes me wonder if something else is going wrong with autofill syncing - eg, maybe we aren't storing metadata correctly so every sync looks like a "first" sync, or something else.

We should consider adding other telemetry to better understand why a fix that we thought shouldn't have a large impact actually did.

┆Issue is synchronized with this Jira Task