[OneEmbedding] Try quantize in Embedding Shuffle (#7912)

* try to add quantize in embedding_shuffle

* add half support

* add numpy quantized test

* Add embedding gradient shuffle quantize

* fix row blockreduce

* use segement quantize and dump right data

* decouple quantize and dequantize kernel'

* add first version of quantize, need test and debug

* fix quantize factor gather error

* fix quantize and dequantize kernel

* remove useless code

* try to add embedding gradient shuffle

* fix shuffle factor size bug

* Add simple env_var control, need refine

* configue the cuda stream

* Add simple warp impl

* Add fully warp impl

* use pack to optimize dequantize kernel

* fix half accuracy

* add more check for embedding shuffle kernel

* Add fully quantize unittest

* simplify code

* Support ComputeType and fix unittest

* fix env name as ONEFLOW_ONE_EMBEDDING_ENABLE_QUANTIZE_COMM

* fix env name

* refine with ComputeType

* fix embedding shuffle unittest

* add round half quantize func

* fix unittest

* fix local test to use different id

* fix jc comment

* fix ONEFLOW_ONE_EMBEDDING_ENABLE_QUANTIZE_COMM to ONEFLOW_ONE_EMBEDDING_ENABLE_QUANTIZED_COMM

* fix name for unittest

* add check for column limit

* add check in enable

* remove unused variable

* add log

* add if def with cuda

* fix compile error

* use std min

Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Juncheng <liujuncheng1022@gmail.com>