Add support bias is none for fused_attention op.

[limin2021]

PR types

Bug fixes

PR changes

OPs

Describe

Fix bus of fused_attention op.
1.Add support for bias is none, as well as corresponding unittest.
2.Add assert for input shape limitations in case of API misuse, where
num_head * head_dim should equal to embed_dim.

Results of related unittests:

[paddle-bot-old]

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

[Superjomn]

请更新具体修改的 Op 定义，每个Op 是否会用到模型推理中
如：

修改 Op定义如下：

1. xx OP, 将 xxattr 改为，会作用到推理
2. xx Op，增加 xxx attr，不会作用到推理

整体结论：修改的Op [会/不会] 影响推理

[limin2021]

所做修改对推理的影响：

1. fused_attention OP中, 将 QKVBias 和LinearBias 的Input attr 改成 .AsDispensable()，不会影响到推理
   整体结论：本次修改 不会影响推理。

训练侧新增的fuse op对推理带来的整体影响，会在后续提一个defusion的pr来专门处理。

[Superjomn]

LGTM

Recently, due to the usage of fusion op in the dygraph program, the jit module will expose the fusion op directlly to inference model, and this will result in the chaotic in the operator granularity in inference model representation.

This requires some new functionality like defusion in jit module, and other members will continue working on this.

Considering this PR is just a bugfix, it is approved.