llama-quant : fix the verification of attention layers for encoder-decoder models

[DamonFool]

llama.cpp fails to quantize T5 models with unequal encoder-decoder blocks.
The failure was caused by the failing of the verification of attention layers.

GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");

The original verification has assumed that the encoder and decoder have the same number of blocks.
So it fails with unequal encoder-decoder models.

Testing: flan-t5-small, t5-small and an unequal encoder-decoder t5 model

[DamonFool]

Hi @CISC , there is another encoder-decoder pr here #16002 .

The simple example is a very good startup to help people get the llama.cpp integrated into their apps.
It would be helpful to also support encoder-decoder models in that example.
Hope you are fine with it
Thanks.

[DamonFool]

Thanks @CISC .