* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* support pure bfloat16

* support bf16 linear

* update PR to pass CI

* tiny fix where_grad_kernel.cu

* Support bfloat16 type for reducer and sharding.

* Fix some bug.

* Polish code.

* Polise code.

* Add bfloat16 datatype in fill_grad kernels.
Co-authored-by: Nsneaxiy <sneaxiy@126.com>
Co-authored-by: Nsneaxiy <sneaxiy@126.com>

* make dense_contents_ lazy init

* update legacy dygraph

* fix legacy dygraph bug

* [CustomDevice] add custom ccl 2/2

* update

* update

* update launch

* remove cudaDeviceContext

* remove more template

* fix rocm compile

* remove alias name CUDADeviceContext

* fix compile

* fix tests

* revert changes

* fix init()

* delete test_device_context

* replace CPUDeviceContext with CPUContext

* fix test_scalar

* remove dot_op.cc

* fix compile

* fix bugs of reducer

* update

* update

* support diff batch for sparse of eagerreducer

* fix

* remove old custom op placetype

* replace dist  placetype using

* add with gpu macro

* fix mutable_data error

* fix set value error

* add comment

* [Dygraph] Support sparse tensor in refactored reducer

* add uts

* refactor

* update

* fix bugs

* rename scalar_array to int_array

* update cmake

* fix conflict

* remove useless log

* replace Backend by Place in C++ API

* fix left code

* fix test_to_api bug

* fix bug

* fix bug

* refactor reducer

* modify cmakelists

* solve conflicts

* rename group and update process_group

* fix bugs of ProcessGroupNCCL

* modify for CIs

* refactoring reducer

* add reducer