fix(backend): skip graph coloring for functions with >250 vregs

For extremely large functions, the iterative spill loop creates a
vreg cascade (176 vregs -> 1675 spilled after 3 rounds) that the
scratch register pool cannot handle. Use all-stack allocation as
a safe fallback for these cases.

Fixes 39_fp_params main/params_mix/params_fa40 output correctness.
params_f40_i24 (176 vregs) and 30_many_dimensions still have
pre-existing computation errors that need the vreg cascade fix.

src/mir/RegAlloc.cpp

@@ -1265,6 +1265,15 @@ namespace mir
       if (function.GetNumVRegs() == 0)
         return;
 
+      // 超大函数跳过图着色，全栈槽分配保证正确性
+      if (function.GetNumVRegs() > 250) {
+        std::set<int> all_spilled;
+        for (size_t i = 0; i < function.GetNumVRegs(); i++)
+          all_spilled.insert(static_cast<int>(i));
+        RewriteWithAllocation(function, {}, {}, all_spilled);
+        return;
+      }
+
       const int MAX_SPILL_ROUNDS = (function.GetNumVRegs() > 120) ? 3 : 10;
       for (int round = 0; round < MAX_SPILL_ROUNDS; ++round)
       {