K2: Postpone inference forks resolution until FULL completion

It helps to use expected type information when choosing the branch
even in the nested calls

^KT-43296 In Progress

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

@@ -0,0 +1,30 @@
+// SKIP_TXT
+
+interface Slice<V>
+
+interface A
+interface B : A
+interface C : A
+
+val SL0: Slice<A> = TODO()
+val SL1: Slice<B> = TODO()
+val SL2: Slice<C> = TODO()
+
+fun <X> foo(s: Slice<X>): X? {
+    if (s.hashCode() == 0) {
+        return id(bar(s))
+    }
+
+    if (s === SL0) {
+        return id(bar(s))
+    }
+
+    if (s === SL1 || s === SL2) {
+        return id(bar(s))
+    }
+    return null
+}
+
+fun <T> id(x: T): T = x
+
+fun <Y> bar(w: Slice<Y>): Y? = null

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

@@ -0,0 +1,30 @@
+// SKIP_TXT
+
+interface Slice<V>
+
+interface A
+interface B : A
+interface C : A
+
+val SL0: Slice<A> = TODO()
+val SL1: Slice<B> = TODO()
+val SL2: Slice<C> = TODO()
+
+fun <X> foo(s: Slice<X>): X? {
+    if (s.hashCode() == 0) {
+        return id(bar(s))
+    }
+
+    if (s === SL0) {
+        return <!TYPE_MISMATCH, TYPE_MISMATCH!>id(bar(<!DEBUG_INFO_SMARTCAST!>s<!>))<!>
+    }
+
+    if (s === SL1 || s === SL2) {
+        return id(bar(s))
+    }
+    return null
+}
+
+fun <T> id(x: T): T = x
+
+fun <Y> bar(w: Slice<Y>): Y? = null