Torch backend: LSTM model training is terribly slow

[DLumi]

LSTM model created with use_cudnn=True is VERY slow comparing to pure-pytorch counterpart on Keras 3.11.3.
Could not test it on higher versions due to #22274
The keras model likely does not truly utilize cuDNN-optimized path as GPU utilization stays really low on training.

Colab to reproduce:
https://colab.research.google.com/drive/1OOjpO4OGNGJi6w0sfxbVv6P7S_-idhXv