Using LoDTensor instead of Tensor in every operator.

[qingqing01]

Fix #4047
Fix #3717

• Use the inheritance in the definition of LoDTensor.
• Modify pybind to expose LoDTensor correctly.
• Replace LoDTensor in every operator.
• Pass unit testing.
  • 我本地很多test_*_op的单测都过了，TeamCity上挂了，待查
• Sequence average pool operator.
  • Just an example.
  • https://github.com/PaddlePaddle/Paddle/pull/4083/files#diff-ca41b976f1749d49345a6eae4c5a62a0
• Share LoD info among each operator.
  • Replace Tensor by LoDTensor in each operator. #4047 (comment)

本地单测都通过：

[build]$ ctest -R _op
Test project /home/dangqingqing/github/myfork/build
      Start   2: parameter_optimizer_test
 1/29 Test   #2: parameter_optimizer_test .........   Passed    0.01 sec
      Start  97: grad_op_builder_test
 2/29 Test  #97: grad_op_builder_test .............   Passed    0.00 sec
      Start 100: net_op_test
 3/29 Test #100: net_op_test ......................   Passed    0.00 sec
      Start 113: test_op
 4/29 Test #113: test_op ..........................   Passed    1.21 sec
      Start 123: test_rowwise_add_op
 5/29 Test #123: test_rowwise_add_op ..............   Passed    6.73 sec
      Start 124: test_uniform_random_op
 6/29 Test #124: test_uniform_random_op ...........   Passed    6.22 sec
      Start 125: test_operator
 7/29 Test #125: test_operator ....................   Passed    1.22 sec
      Start 126: test_mul_op
 8/29 Test #126: test_mul_op ......................   Passed   58.44 sec
      Start 127: test_add_two_op
 9/29 Test #127: test_add_two_op ..................   Passed    5.43 sec
      Start 128: test_recurrent_op
10/29 Test #128: test_recurrent_op ................   Passed    5.07 sec
      Start 129: test_scale_and_identity_op
11/29 Test #129: test_scale_and_identity_op .......   Passed    4.88 sec
      Start 132: test_cross_entropy_op
12/29 Test #132: test_cross_entropy_op ............   Passed   10.67 sec
      Start 133: test_squared_l2_distance_op
13/29 Test #133: test_squared_l2_distance_op ......   Passed    6.08 sec
      Start 134: test_reshape_op
14/29 Test #134: test_reshape_op ..................   Passed    7.61 sec
      Start 135: test_cos_sim_op
15/29 Test #135: test_cos_sim_op ..................   Passed    5.91 sec
      Start 137: test_elementwise_mul_op
16/29 Test #137: test_elementwise_mul_op ..........   Passed    7.41 sec
      Start 139: test_pad_op
17/29 Test #139: test_pad_op ......................   Passed    4.99 sec
      Start 140: test_gather_op
18/29 Test #140: test_gather_op ...................   Passed    4.72 sec
      Start 141: test_minus_op
19/29 Test #141: test_minus_op ....................   Passed    9.49 sec
      Start 143: test_mean_op
20/29 Test #143: test_mean_op .....................   Passed    6.11 sec
      Start 144: test_gaussian_random_op
21/29 Test #144: test_gaussian_random_op ..........   Passed    5.68 sec
      Start 145: test_sigmoid_op
22/29 Test #145: test_sigmoid_op ..................   Passed    5.01 sec
      Start 147: test_concat_op
23/29 Test #147: test_concat_op ...................   Passed    3.68 sec
      Start 148: test_sgd_op
24/29 Test #148: test_sgd_op ......................   Passed    5.08 sec
      Start 150: test_top_k_op
25/29 Test #150: test_top_k_op ....................   Passed    1.27 sec
      Start 151: test_softmax_op
26/29 Test #151: test_softmax_op ..................   Passed    6.23 sec
      Start 153: test_scatter_op
27/29 Test #153: test_scatter_op ..................   Passed    3.79 sec
      Start 154: test_sum_op
28/29 Test #154: test_sum_op ......................   Passed    6.66 sec
      Start 155: test_fill_zeros_like_op
29/29 Test #155: test_fill_zeros_like_op ..........   Passed    5.97 sec

100% tests passed, 0 tests failed out of 29

Total Test time (real) = 195.61 sec

[qingqing01]

Specialized template function InferShapeContext::Input and InferShapeContext::MultiInput, thus, the InferShape of non-sequence operators can call Input<Tensor> to avoid to use LoDTensor.

[wangkuiyi]

I think this specialization is great to hide the difference between LoDTensor and Tensor. I can also accept that we don't have this and expose LoDTensor to non-sequence operator contributors, because I feel that LoDTensor is a key difference from TensorFlow and could be a PaddlePaddle flag.

[qingqing01]

Redifined and specialized template function ExecutionContext::Output and ExecutionContext::MultiOutput, thus, the kernels of non-sequence operators can call Input<Tensor> and Output<Tensor> to avoid to use LoDTensor.

[wangkuiyi]

LGTM++

[qingqing01]

How to share LoD info in operator will be done in next PR, thus merged this PR.