codegen: make collection aliases (Map/List/Set/…) type-check and construct

[code-by-sia]

Two coupled bugs made a collection alias (type M = Map<K, V>, List<T>, Set<T>, Stack<T>, …) unusable, and the second one was a silent runtime crash.

Bugs

1. Typedef ordering — genRefinedTypedefs emitted typedef xc_Map_string_User_t xc_M_t; before genArrTypedefs defined the underlying xc_Map_string_User_t, so a compound value gave unknown type name.
2. Silent null-struct — empty M emitted a zeroed (xc_M_t){0} (a null map/list with no backing storage) that compiled fine and then crashed at first .put/.get. Fixing (1) alone would have turned the compile error into this silent runtime crash, so they had to be fixed together.

Fix

1. isCompositeAlias now defers any collection-ctype alias (xc_Map_, xc_List_, xc_Set_, …) to genAliasTypedefs, which runs after the monomorphization typedefs exist — exactly how xc_arr_ aliases were already handled.
2. genPrimary resolves a bare empty <collection-alias> to the real runtime constructor (xstd_map_new / xstd_list_new / xstd_set_new / …), reconstructing the element ctypes from the alias base type (the key suffix is always a single known word, so the split is unambiguous).

type UserMap = Map<String, User>
let m = empty UserMap           // now a real map, was a null struct
m.put("ada", User { name: "Ada", age: 36 })

Verification

• Regression test examples/language/collection_alias_test.xi (Map with compound + primitive values, List, Set).
• Probed edge cases: alias as class state, Stack/Set aliases, nested Map<String, List<Integer>> — all work.
• Self-host fixpoint holds (gen2 == gen3); all 46 existing test files pass.