This PR made these changes to support double grad:

1. Translate `paddle.grad` to `paddle.static.gradients` to support double grad for dy2stat.
2. Fix IfElseTransformer bug which may not change value if "Store before Load" variable is in "Store" statement is in IfElse conditional statement
3. Add `DOut` to support double grad variables in `run_program_op`
4. Add support for renaming for double grads for `jit.save/load`

* Dygraph Recompute support AMP

* dygraph recompute: update unitest

* fix the bug of div operation result using scale method do not exactly equal the result of elementwise_div method

* remove __div__ , __rdiv__ methods which do not define in python3

* modify the note

* add test case

* add test case

* add trt convert op: reshape

* reshape

* fix

* input: Shape,ShapeTensor

* Add some comments

* fix spawn default nprocs get error

* polish error message

* fix syncbn

* enhance error message for conv

* fix ci coverage

* improve group norm cpu precision and performance

* add unit test to group norm

* support params groups, test=develop

* simplify updating opt attr

* update according to review

* fix uint8 check in assign

* fix assign unittests

* fix xpu test_assign

* add more conditions

* dynamic shape

* ut

* correct contidions

* commnent

* remove rebandadnt op type

* remove rebandant if

* run pre-commit

* use HOST or DEVICE instead of HOSTDEVICE in implementation of lgamma op

* add test for fp32

* add lgamma to op_threshold_white_list

* add cuda kernel for lgamma kernel

* modify numeric grad delta

* fix small English issue

* change LaunchElementwiseCudaKernel to LaunchSameDimsElementwiseCudaKernel

Change-Id: I2e82dfcee6a1d0512b94cebc32281123fa5bf597

* pretty print for datafeed error

Change-Id: I056a8b6f03608e96679a83846c97aed289cef7e6

* fix fleet dist infer ut

* refine ~npuOpRunner

* implement destructor and forbid copy

* use reference to avoid copy

* use const reference

* relax adam precision

* fix top_k

Sigmoid: Out = Sigmoid(X)
SigmoidGrad: DX = DOut*(1-Out)*Out

[This Patch]
Out
DOut -> SigmoidGradGrad -> DOutNew
DDX                        DDOut

DDOut = (1-Out)*Out*DDX
DOutNew = (1-2*Out)*DOut*DDX

* added op cast functionality for fp32/bf16

* added newline

* added entries in static mode white list and unity build

* fixed failing tests

* changes after review

* added formatting

* upgraded tests file as reviewer suggested

* changes after review

* minor change

* fix pipeline

* fix mp pp dp

* fix utest of hybrid parallel

* add utest for tuple

* added support for most matmul cases

* added more functionality

* full functionality of matmul op, fp32 only

* added bf16 tests and functionality

* added formatting

* changes after review

* minor change

* added reviewers suggestions

* replace complex in set tensor from and to numpy

* replace complex template in cast op

* cast indices and label if their type is not consistent

* fix bug

* add unittest

* fix gather op and add logsumexp op on kunlun

* update xpu depence

* update tests and fix elementwise_add

* Support convert sublayers in Sequential Container

* remove paddle.jit.set_code_level

* fix test_paddle_save_load and test_paddle_save_load_binary

* fix unittest:test_paddle_save_load and test_paddle_save_load_binary

* delete *.pyc

* add comment for unittest