fix fused_qkv model accuracy issue (microsoft#5217)

Fused_qkv model can not correctly choose the fused_qkv type. Need to
update the module_name_matches.

Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>

deepspeed/module_inject/fusedqkv_utils.py

@@ -113,7 +113,7 @@ def _transpose_fused_qkvw(src, mp_size, fused_qkv_type=None, module=None):
 
         raise ValueError("unknown fused_qkv_type")
 
-    module_name_matches = [k for k in fused_type_dict.keys() if module_str in k]
+    module_name_matches = [k for k in fused_type_dict.keys() if k in module_str]
     if module_name_matches:
         # There can be overlap with matches (e.g., "DecoderLayer" and "FalconDecoderLayer").
         # We take the longest matching module_name