frontend: make fn calls byval; fix false positive isNonErr

This commit does two things which seem unrelated at first, but,
together, solve a miscompilation, and potentially slightly speed up
compiler perf, at the expense of making #2765 trickier to implement in
the future.

Sema: avoid returning a false positive for whether an inferred error set
is comptime-known to be empty.

AstGen: mark function calls as not being interested in a result
location. This prevents the test case "ret_ptr doesn't cause own
inferred error set to be resolved" from being regressed. If we want to
accept and implement #2765 in the future, it will require solving this
problem a different way, but the principle of YAGNI tells us to go ahead
with this change.

Old ZIR looks like this:

  %97 = ret_ptr()
  %101 = store_node(%97, %100)
  %102 = load(%97)
  %103 = ret_is_non_err(%102)

New ZIR looks like this:

  %97 = ret_type()
  %101 = as_node(%97, %100)
  %102 = ret_is_non_err(%101)

closes #15669

test/behavior/error.zig

@@ -938,3 +938,28 @@ test "returning an error union containing a type with no runtime bits" {
     var zero_byte: ZeroByteType = undefined;
     (&zero_byte).* = try ZeroByteType.init();
 }
+
+test "try used in recursive function with inferred error set" {
+    if (builtin.zig_backend == .stage2_aarch64) return error.SkipZigTest; // TODO
+    if (builtin.zig_backend == .stage2_arm) return error.SkipZigTest; // TODO
+
+    const Value = union(enum) {
+        values: []const @This(),
+        b,
+
+        fn x(value: @This()) !void {
+            switch (value.values[0]) {
+                .values => return try x(value.values[0]),
+                .b => return error.a,
+            }
+        }
+    };
+    const a = Value{
+        .values = &[1]Value{
+            .{
+                .values = &[1]Value{.{ .b = {} }},
+            },
+        },
+    };
+    try expectError(error.a, Value.x(a));
+}