[region-isolation] Reword merge-region diagnostics so we never fail on missing names

[gottesmm]

Restructure the five "incompatible region merge" diagnostics emitted by IncompatibleRegionMergeDiagnosticEmitter so that the primary error never depends on recovering a value name. Previously, when VariableNameInferrer could not infer a name for either side of the merge, every emitter fell back to emitUnknownPatternError() and the user got the unhelpful "we do not understand this pattern" diagnostic for a violation the analysis had in fact understood.

Now:

• Each primary states the operation that performs the merge and the two isolation domains involved -- e.g. "passing arguments to %kind2 could allow for references between values exposed to %0 and %1 code, risking data races". The primary requires only the isolation strings and (where applicable) the callee decl or cast target type.

• Per-value details move to follow-up notes anchored at each value's defining source location (parameter decl for SILFunctionArgument, defining-instruction loc otherwise) so the user can see where each side came from. Two note variants: named ("'self.x' is exposed to ... code") when name inference succeeds, bare ("value is exposed to ... code") otherwise. We deliberately do not emit a "value of type T" variant because SIL values often do not carry a reliable AST type.

• The shared emitMergeRegionValueNote() helper centralises note emission and the new preciseSourceLocForValue() helper picks the most precise SourceLoc available for a value.

emitNonisolatedFunction now falls through to emitUnknown rather than emitUnknownPatternError when ApplySite/getCalleeDeclRef recovery fails, so the primary still fires with isolation information in those cases. emitUnknownPatternError remains reachable only when isolation info itself is invalid -- a genuinely unknown pattern.

rdar://175592408

[gottesmm]

@swift-ci smoke test