Cache final impl lookup results. (#6452)

If an impl lookup finds a final result, cache that and reuse it if we
perform the same lookup later.

In addition to reducing repeated work, this allows us to produce the
same result for repeated lookups that find a C++ operator. This isn't a
great solution to that problem, as it's not clear how to extend it to
behave correctly across import, but we don't have a solution for that
for C++ interop in general.

toolchain/check/cpp/impl_lookup.cpp

@@ -101,6 +101,9 @@ static auto BuildWitness(Context& context, SemIR::LocId loc_id,
         if (struct_value.type_id == SemIR::ErrorInst::TypeId) {
           return SemIR::ErrorInst::InstId;
         }
+        // TODO: If a thunk is needed, this will build a different value each
+        // time it's called, so we won't properly deduplicate repeated
+        // witnesses.
         witness_value_id = CheckAssociatedFunctionImplementation(
             context,
             context.types().GetAs<SemIR::FunctionType>(struct_value.type_id),