optimize compilation of operators using eigen

[Avin0323]

PR types

Others

PR changes

Others

Describe

优化使用Eigen库broadcast调用的部分op：

1. 新增paddle/fluid/operators/eigen目录，与eigen相关的代码在此目录下实现；
2. 算子实现中，使用统一封装的函数来替换y.device(dev) = x.xxx();形式；
3. 本PR只涉及broadcast调用，其他高频函数后续逐步优化；

• 库体积变化：

单架构下预测静态库减少14M(644M->630M)，动态库体积减少6M(244M->238M)，Whl包体积减少1M(115M->114M)

• 编译耗时变化：

*.o	耗时变化	变化比例
expand_as_op.cu.o	56s -> 37s	降低33.9%
expand_as_op.cc.o	56s -> 23s	降低58.9%
expand_v2_op.cu.o	89s -> 60s	降低32.5%
expand_v2_op.cc.o	91s -> 57s	降低37.3%
meshgrid_op.cu.o	58s -> 38s	降低34.4%
meshgrid_op.cc.o	48s -> 21s	降低56.2%
tile_op.cu.o	85s -> 57s	降低32.9%
tile_op.cc.o	90s -> 56s	降低37.7%
expand_op.cu.o	85s -> 55s	降低35.2%
expand_op.cc.o	89s -> 53s	降低40.4%
expand_as_v2_op.cu.o	58s -> 39s	降低32.7%
expand_as_v2_op.cc.o	61s -> 25s	降低59%
addmm_op.cu.o	39s -> 39s	持平
addmm_op.cc.o	23s -> 21s	降低8.6%

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[wanghuancoder]

LGTM

[zhangting2020]

LGTM