Fix sticky diagnostics

[pepeiborra]

In haskell/ghcide#959 I tried to fix a problem with diagnostics:

• PROBLEM: when early cut-off kicks in and a stage is skipped (e.g. TypeCheck) its diagnostics are lost
• SOLUTION: when creating diagnostics for a new version of a document, start from the diagnostics for the previous version.

The assumption in the solution is that starting with the previous diagnostics is sound because they will get updated incrementally as the rules are re-run. If a rule is not re-run then it must have been cut-off and its diagnostics are still valid.

But hold on, the "build" graph is dynamic, so the rules to be run for a file can change at any time! So it's not always the case that if a rule is not re-run, its diagnostics are still valid. For instance, the graph for a non-FOI changes when it becomes a FOI, e.g. GetModIfaceFromDisk is not re-run, so its diagnostics will stick around until the file becomes a non-FOI again, and GetModIfaceFromDisk is re-run.

The right solution is to store the diagnostics in our state (the values ref):

• when creating diagnostics for a new version, just start from scratch as we used to
• when early cut-off kicks in, reuse the diagnostics stored during the previous run of the rule
• graph changes are not a problem, since only the diagnostics from the rules that actually run are considered

Fixes #1171