Add support for INT8 quantization of fusion_gru op

[wojtuss]

Please, enable INT8 quantization of the fusion_gru op using oneDNN-based INT8 kernel.

[wojtuss]

After enabling INT8 quantization of the fusion_gru op (a PR will be submitted soon) we achieved the following accuracy result for the QAT GRU model on a CLX 6248 (with VNNI support):

	qat (fp32)	int8	diff
Precision	0.89198	0.89221	0.00023
Recall	0.89449	0.89412	-0.00037
F1 score	0.89323	0.89316	-0.00007

[lidanqing-intel]

PR #27481 submitted

[wojtuss]

A support for INT8 quantization of the GRU model is added to Paddle. The only piece we are waiting to add here is a version of oneDNN with additional GRU INT8 optimizations. The upgrade of oneDNN will be submitted in the PR #28420.

[wojtuss]

The PR #28420 is merged so the support for INT8 quantization of the GRU model is done.

[paddle-bot-old]

Are you satisfied with the resolution of your issue?

YES
No