[Bugfix] Remove redundant T.fill to fix precision issue (#667)

Author: xuchangtolearn

examples/flash_decoding/example_gqa_decode.py

@@ -169,7 +169,7 @@ def flash_attn_split(
             T.fill(scores_max, -T.infinity(accum_dtype))
 
             loop_range = T.ceildiv((seqlen_kv // num_split), block_N)
-            T.fill(K_shared, 0)
+
             for k in T.Pipelined(loop_range, num_stages=num_stages):
                 T.copy(
                     K[bid, (seqlen_kv // num_split) * sid +