Fix full_op bug for gradient merge

[From00]

PR Category

Auto Parallel

PR Types

Bug fixes

Description

auto_parallel_gradient_merge pass中通过full算子输出scale_op的scalar参数，在full op中未将scalar输出数据指定为cpu place，导致实际运行时scalar参数产生冗余的h2d拷贝，并可能引起pipeline parallel 1F1B调度hang住。本PR修复此问题。

Pcard-76459

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}