gs (#48288)

* add granite speech 5

* nits

* add pipeline snippet

* more nits

* yet another nit

* no need

* fullgraph compile works

* add device in AudioKwargs

* faster FE

* update doc

* clarify config.num_mel_bins * 4

* inherit GraniteSpeech5EncoderAttention

* nit

* remove get_num_encoder_frames

* remove einsum

* add keep in fp32 strict

* conversion script update

* fixtures reproducer

* make

* TRF038 does not apply here