[Cherry-pick 1.6] fix batch_norm_grad shape=0 & allreduce shape enforce & sync_batch_norm hang in fleet (#22157)

* Fix softmax cuda bug

* Refine multihead log and softmax logic

* Align block to 32

* fix seqconv_eltadd_relu pass during multi-threads predictor, test=develop

* fix attention_lstm_fuse_pass during multi-threads inference, test=develop

* fix embedding_fc_lstm_fuse_pass during multi-threads inference, test=develop

* fix fc_lstm_fuse_pass during multi-threads inference, test=develop

* fix seq_concat_fc_fuse_pass during multi-threads inference, test=develop

* add param & grad shape check for sgd op

* add _reshape_inplece interface for dygraph parallel

* refine unittest based paddle/models scripts, test=develop

* add unittest for parallel grad fuse, test=develop

* fix decay param in DecayAdagrad test=develop (#22026)

* fix integer overflow in match_matrix (#22036)

* fix integer overflow in match_matrix test=develop

* fix integer overflow in match_matrix test=develop

* fix typo test=develop

This reverts commit 0473cdb8.

test=release/1.6

* fix the computation for dx (grad for x) for prelu operation. (#20949)

* set the default value of alpha for prelu to 0.25, test=develop

* add the call to __syncthreads(), test=develop

* fix the implementation of cpu prelu, test=develop

* repair the implementation of element mode prelu, test=develop

* modify test_prelu_op.py, test=develop

make config option DisableGlogInfo() able to mute all inference logs

* fix fetch handler problem and refactor
when a user define FetchHandler class, he or she should initialize a handler
with variable dict. the key of a variable dict is a user defined name,
the value of a variable dict is a Varaible generated from python API.

For each fetching, a user should implement handler function in which
fetched_result_dict will be available and the user can access the fetched value
with user defined keys.

* ADD NV JETSON SUPPORT
test=release/1.6

* CHERRY_PICK: specify the auto growth allocator for inference.
test=release/1.6

* disable reshape inplace in dygraph model; test=develop (#21157)

* fix ExecutionContext::HasInput and ExecutionContext::HasOutput depend on the scope structure, test=develop (#20721)

* set dim[0] to -1 if dim[0] < 0 and remove assertion to runtime, test=develop

* fix transpose conv,test=develop

* fix comments
test=develop

* Improve argsort performance.

- Give 200000 data to compute argsort on v100,
can speed up ~190x
before opt cost: 0.53s
after opt cost:0.0027s

- Add fp16 support

* Refine error message
* Refine code
* Add descending sort

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* add Adam beta1/beta2 support Variable. test=develop

* Add Asypadding for conv fusion.

test=develop

reference: pr/20042

* Fix eigen build link error

* Change back file mode

* Use math function & add more checks.

* add the framework support for distfc and ut, test=develop
* fix the implementation of shard_index_op, test=develop

* batch_norm momentum support variable. test=develop

* support dump param of model into afs (#20302)

* support dump param to afs
test=develop

* code style
test=develop

* code style
test=develop

* dump param
test=develop

* dump param
test=develop

* dump param
test=develop

* dump param
test=develop

* find lookup table in order (#20932)

test=develop

* cherry-pick
test=develop

* solve pslib core in stop worker
test=develop

* print table stat info for pslib
test=develop

* Improve topk performance.

give 200000 data to compute topk,
before opt: cost 1s
after opt: cost 0.0028s.

* Refine return value.
* Add cuda util funtions.
* Fix ComputeBlockSize bug & refine comments.
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

The op should handle k=1024
Fix seq_len < warpsize error.

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

seems shuffle_sync cannot handle small size

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* Fix the crash issue when scale or bias was null-pointer.

* Add the error message for passing CI.

test=release/1.6

added splitter "__" between weight name and suffix number to avoid conflicts.