Use first match branch in case of type mismatch, not last

The comment says this was intentional, but I do agree with #4304 that it makes
more sense the other way around (for if/else as well).

Fixes #4304.

Author: flodiebold

crates/ra_hir_ty/src/infer/coerce.rs

@@ -45,9 +45,7 @@ impl<'a> InferenceContext<'a> {
                 self.coerce_merge_branch(&ptr_ty1, &ptr_ty2)
             } else {
                 mark::hit!(coerce_merge_fail_fallback);
-                // For incompatible types, we use the latter one as result
-                // to be better recovery for `if` without `else`.
-                ty2.clone()
+                ty1.clone()
             }
         }
     }

crates/ra_hir_ty/src/tests/simple.rs

@@ -937,7 +937,7 @@ fn main(foo: Foo) {
     51..107 'if tru...     }': ()
     54..58 'true': bool
     59..67 '{      }': ()
-    73..107 'if fal...     }': ()
+    73..107 'if fal...     }': i32
     76..81 'false': bool
     82..107 '{     ...     }': i32
     92..95 'foo': Foo