Tensor type promotion for static mode

[zoooo0820]

PR types

New features

PR changes

APIs

Description

Pcard-78266

This PR is based on #59513 and #59518

support FLOAT type promotion of Tensor-Tensor for following operaters: +(elementwise_add) / -(elementwise_sub) / *(elementwise_mul) in static mode.

From this PR these operators will meet commutative law (currently, only for FLOAT type)，instead of align to former tensor. For example:

>>> paddle.enable_static()
>>> a = paddle.randn((2,3), dtype='float32')
>>> b = paddle.randn((2,3), dtype='float64')
>>> out = a + b

# Before, dtype of out would be same to the former one, i.e. a
>>> out
var tmp_1 : LOD_TENSOR.shape(2, 3).dtype(float32).stop_gradient(True)


# Now, dtype will be promote to float64
>>> out
var tmp_1 : LOD_TENSOR.shape(2, 3).dtype(float64).stop_gradient(True)

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zxcd]

LGTM

[jeff41404]

LGTM