fix: rope's weights sometime need to hold in lmem

- if rope's weight don't need to reload, add hold_in_lmem tag

Change-Id: I43099e0d32fb58fd7985ae42b214dbd49d2b141f

Author: yifan.chen

lib/Dialect/Tpu/Transforms/LayerGroup/BasicTimeStep.cpp

@@ -353,7 +353,7 @@ void BasicTimeStep::gen_hold_coeff() {
           if (src_op.getAllowSplitAttr() != nullptr) {
             allow_split = true;
           }
-          if (allow_split == false) {
+          if (allow_split == false || tensor_info.hold_in_lmem) {
             this->hold_coeff_[v] = ts;
           }
         }

lib/Dialect/Tpu/Transforms/LayerGroup/LayerGroupUtil.cpp

@@ -1618,6 +1618,7 @@ bool get_backward_slice_info(LgInfo &lg_info, slice_info_t &in_si,
         idx = 0;
         slice = 1;
         if (!rope_need_reload) {
+          hold_in_lmem = true;
           in_si.n.emplace_back(slice_pair_t(idx, slice));
           break;
         }
@@ -1752,6 +1753,7 @@ bool get_backward_slice_info(LgInfo &lg_info, slice_info_t &in_si,
         idx = 0;
         slice = 1;
         if (!rope_need_reload) {
+          hold_in_lmem = true;
           in_si.h.emplace_back(slice_pair_t(idx, slice));
           break;
         }