NI: Fix regression for star-projections approximation

See the test added

^KT-35703 Fixed

compiler/testData/diagnostics/tests/inference/starApproximationFromDifferentTypeParameter.fir.kt

@@ -0,0 +1,20 @@
+// SKIP_TXT
+// !LANGUAGE: +NewInference
+
+// See KT-14453 and KT-35703
+val <T : Any> KClass1<T>.primaryConstructor: KFunction1<T>? get() = null!!
+
+interface KClass1<F : Any>
+interface KFunction1<out R>
+fun f(type: KClass1<*>): KFunction1<Any>? =
+    // What happens here:
+    // 1. Create captured type for type argument T (it's built upon star-projection from `KClass<*>` that's argument is <: Any)
+    // 2. Approximate resulting descriptor, now it 'KClass1<*>.primaryConstructor: KFunction1<*>'
+    //    Note that star-projection in 'KFunction1<*>' is obtained from KClass and its `projectionType` is Any (not-nullable).
+    //    Of course that situation is already strange
+    // 3. When checking subtyping after call completion we get that 'KFunction1<*>' is not a subtype of 'KFunction1<Any>' in new type checker.
+    //    The answer is correct, but this old type checker used `projectionType` for this and it was resulting in KFunction1<*> <: KFunction1<Any>
+    //
+    // So to fix the issue we use 'out starProjectionType' instead of star-projection itself in approximation
+    // Note that in new inference there should be no approximation for IN-position types (they must remain captured)
+    type.primaryConstructor

compiler/testData/diagnostics/tests/inference/starApproximationFromDifferentTypeParameter.kt

@@ -0,0 +1,20 @@
+// SKIP_TXT
+// !LANGUAGE: +NewInference
+
+// See KT-14453 and KT-35703
+val <T : Any> KClass1<T>.primaryConstructor: KFunction1<T>? get() = null!!
+
+interface KClass1<F : Any>
+interface KFunction1<out R>
+fun f(type: KClass1<*>): KFunction1<Any>? =
+    // What happens here:
+    // 1. Create captured type for type argument T (it's built upon star-projection from `KClass<*>` that's argument is <: Any)
+    // 2. Approximate resulting descriptor, now it 'KClass1<*>.primaryConstructor: KFunction1<*>'
+    //    Note that star-projection in 'KFunction1<*>' is obtained from KClass and its `projectionType` is Any (not-nullable).
+    //    Of course that situation is already strange
+    // 3. When checking subtyping after call completion we get that 'KFunction1<*>' is not a subtype of 'KFunction1<Any>' in new type checker.
+    //    The answer is correct, but this old type checker used `projectionType` for this and it was resulting in KFunction1<*> <: KFunction1<Any>
+    //
+    // So to fix the issue we use 'out starProjectionType' instead of star-projection itself in approximation
+    // Note that in new inference there should be no approximation for IN-position types (they must remain captured)
+    type.primaryConstructor