Fix performance regression in NI by fixing totally incorrect hashCode

The actual problem was introduced in 4f1e85b468, note how `hashCode` is implemented:
```
var currentHashCode = cachedHashCode
if (currentHashCode == 0) return currentHashCode
...
```
It's a silly bug, there should be check `if (currentHashCode != 0) ...` because `0` is used a marker for "uncomputed value".

Now, in the commit 0219b86d06 I added map with `KotlinType` as a key and because of constant `hash` for `KotlinType`, we basically got `List` instead of `Map`, which caused this performance regression

 #KT-34063 Fixed

compiler/tests/org/jetbrains/kotlin/checkers/DiagnosticsTestGenerated.java

@@ -10131,6 +10131,11 @@ public class DiagnosticsTestGenerated extends AbstractDiagnosticsTest {
                     runTest("compiler/testData/diagnostics/tests/inference/commonSystem/lessSpecificTypeForArgumentCallWithExactAnnotation_ni.kt");
                 }
 
+                @TestMetadata("manyArgumentsForVararg.kt")
+                public void testManyArgumentsForVararg() throws Exception {
+                    runTest("compiler/testData/diagnostics/tests/inference/commonSystem/manyArgumentsForVararg.kt");
+                }
+
                 @TestMetadata("nestedLambdas.kt")
                 public void testNestedLambdas() throws Exception {
                     runTest("compiler/testData/diagnostics/tests/inference/commonSystem/nestedLambdas.kt");