refine infershape of sequence_enumerate, hash and fuse_emb_seq_pool

[luotao1]

• move SetOutputDim and ShareLoD of Infershape's Runtime into Compute function.
• test PyramidDnn inference:
  • command: ./paddle/fluid/inference/tests/api/test_analyzer_pyramid_dnn --infer_model=third_party/inference_demo/pyramid_dnn/model/ --infer_data=third_party/inference_demo/pyramid_dnn/data.txt --gtest_filter=Analyzer_Pyramid_DNN.profile -repeat=1000
  • machine: 2620 v3

Speedup: 0.272498ms->0.222548ms, 18%.

• before:

Event                                Calls       Total       Min.        Max.        Ave.        Ratio.
thread0::fused_embedding_seq_pool    16000       83.6798     0.004181    0.066233    0.00522999  0.287419
thread0::hash                        12000       83.5048     0.005935    0.21865     0.00695873  0.286818
thread0::sequence_enumerate          12000       44.1801     0.003242    0.358377    0.00368168  0.151748
thread0::fc                          2000        37.0386     0.01693     0.202705    0.0185193   0.127218
...

• after

Event                                Calls       Total       Min.        Max.        Ave.        Ratio.
thread0::hash                        12000       72.19       0.004968    0.07149     0.00601583  0.289498
thread0::fused_embedding_seq_pool    16000       64.7578     0.002993    0.34768     0.00404736  0.259694
thread0::fc                          2000        37.7587     0.017463    0.130483    0.0188793   0.151421
thread0::sequence_enumerate          12000       31.06       0.002213    0.105502    0.00258833  0.124558
...

[tensor-tang]

LGTM

[tensor-tang]

lod如果这个op没用，感觉可以尽量早点set比较好

[luotao1]

好的，会在下一个PR修改。