Re-reverting 8274

tir.allocate should be lowered to tvm built-in
in order for them to lowered TVMBackendAllocWorkspace
calls and not forced onto the stack.

Change-Id: I393b3c4580540d699e8ec3e4d62bb5f8c479553e

src/tir/transforms/lower_tvm_builtin.cc

@@ -113,16 +113,6 @@ class BuiltinLower : public StmtExprMutator {
     op = stmt.as<AllocateNode>();
     // Get constant allocation bound.
     int64_t nbytes = GetVectorBytes(op->dtype);
-    if (device_type_.defined()) {
-      if (const auto* dev_type = device_type_.as<IntImmNode>()) {
-        if (dev_type->value == kDLCPU) {
-          int32_t constant_size = op->constant_allocation_size();
-          if (constant_size > 0 && constant_size * nbytes < runtime::kMaxStackAlloca) {
-            return stmt;
-          }
-        }
-      }
-    }
     PrimExpr total_bytes = make_const(op->extents[0].dtype(), nbytes);
     for (size_t i = 0; i < op->extents.size(); ++i) {
       total_bytes = total_bytes * op->extents[i];