[cherry-pick][hybrid performance] Grad fuse for gradient merge under pipeline mode (#35004) (#35299)

* enable npu alignment

* support flatten_params/grads

* support clip by global norm

* remove memset in coalesce_tensor_op

* fix npu kernel of sum op when input is one tensor

* add ut for flatten_param_grads+regularizer

* fix ut

* fix typo

test=develop
test=op_version

* Incorporate cudnn_lstm into LSTM api.
test=develop

* Make coalesce_tensor support alignment optionally.
test=develop

* Reorganize RNN apis. test=develop

* Fix cudnn rnn layout conversion.
test=develop

* Add sequence_length support for RNN cudnn implement.
Add optional init_h and init_c gradient for cudnn_lstm_op.
test=develop

* Use create_parameter for rnn cudnn impl.
test=develop

* Move `self._flat_weight = self.create_parameter()` in RNNBase to main_program.
test=develop

* Update RNN api unittest to use set_device.
test=develop

* Fix set_place for unit tests of RNN apis.
test=develop

* Fix use_align in coalesce_tensor_op.
test=develop

* Adjust RNN apis arguments according to comments.
test=develop

* Polish documents for SimpleRNN apis.
test=develop

* Refine random seed in cudnn_lstm_op.
Expose rnn params from sublayers to RNN.
test=develop

* Fix RNN saving for jit.save.
Refine cudnn_lstm dropout behavior.
test=develop

* Fix doc of GRU. test=develop

* Use ShareDataWith to avoid copying for cudnn_lstm_op test.
test=develop

* Remove updates on cudnn_lstm temporarily.
test=develop

* Use ShareDataWith to avoid copying for cudnn_lstm_op test.
test=develop

* Refine random seed in cudnn_lstm_op.
test=develop

* Fix test_lstm by adjust ConcreteProgram buffer getter.
test=develop

* Use create_parameter instead of create_var for rnn._flat_weight for static graph usage.
test=develop

* Remove W input for cudnn_lstm to pass unused_var_check.
test=develop

* Add test_predict for RNN unit tests coverage.
test=develop

* Fix code style of rnn.
test=develop

* Fix F.rnn usage in rnn.py.
test=develop

* polish no onwer op error message

* fix unittest failed

* polish details based reviewer comment

* Refine adam op, test=develop

* Fuse kernels together to reduce cpu time.

* Refine paddle enforce, test=develop

* Remove some comments, test=develop

* Refine code,test=develop

* Refine cuda kernel, test=develop

* Refine code according to comments, test=develop

* add_dygraph_execution_context

* add dygraph infershape context and execution context; test=develop

* fix imperative bug; test=develop

* remove inputs outputs interface from execution context,
because it have same function with inputNames;
test=develop

* remove tracer_test ctest; test=develop

* fix split op bug; test=develop

* fix unitests bug; test=develop

* fix distribute test bug; test=develop

* fix ngraph compile bug; test=develop

* fix grad maker bug; test=develop

* fix load op bugs; test=develop

* fix operator.cc construct bug; test=develop

* remove useless name find in operator; test=develop

* add tracer_test; test=develop

* fix concat, split bug; test=develop

* remove tracer_test unitest; test=develop

* fix attribute check bug; test=develop

* add test code to fix converage; test=develop

* remove useless code, change check backward input in engin; test=develop

* unlock var type infer shape;test=develop

* add ShareAllLoD api; test=develop

* add dygraph infershape context unitest; test=develop

* remove increase and decrease lod in dygraph; test=develop

* addd override; test=develop

* fix increase descrease lod; test=develop

* fix paddle_enforce; test=develop

* disable lod op dygraph check; test=develop

* fix paddle enforce error; test=develop

* add comment for op_registry and OperatorBase; test=develop

* optimize the comment of op_registry; test=develop

* fix format of comment; test=develop

* fix format of comment; test=develop

* optimize the format of comment; test=develop

* optimize the format of the comment; test=develop

* optimize comment of op_registry; test=develop

* add pre condition check for fuse optimizer op pass, test=develop

* add log & set init to zero, test=develop

* fix test_fuse_all_reduce_pass failed, test=develop

* polish details, test=develop

* refine PADDLE_ENFORCE & remove needless VLOG, test=develop

* refactor op check method, test=develop

Add dtype for coalesce_tensor_op

* support sparse gradients
test=develop

* fuse optimizer

* add alloc_continuous_space_op
test=develop

* Polish code
test=develop

* follow comment
test=develop