compiler: avoid zero-sized alloca in channel operations

This works around a bug in LLVM
(https://bugs.llvm.org/show_bug.cgi?id=49916) but seems like a good
change in general.

compiler/compiler_test.go

@@ -51,6 +51,7 @@ func TestCompiler(t *testing.T) {
 		{"pragma.go", ""},
 		{"goroutine.go", "wasm"},
 		{"goroutine.go", "cortex-m-qemu"},
+		{"channel.go", ""},
 		{"intrinsics.go", "cortex-m-qemu"},
 		{"intrinsics.go", "wasm"},
 	}