[generate] handle support for cache classes when num enc layers != num dec layers

[gante]

What does this PR do?

I think this bug has been in since EncoderDecoderCache was added: we naively assumed that # encoder layers == # decoder layers, as we used config.num_hidden_layers (# encoder layers in encoder-decoder models) to instantiate both caches 👀

This PR adds the missing logic:

• we can specify which part of the config we want to pull with get_text_config() (previously we could only isolate the decoder). We also update num_hidden_layers accordingly, so that the caches see the right number of layers.
• we pull the right config (encoder vs decoder) before parameterizing EncoderDecoderCache in generate
• (adds tests)

Fixes #40120

---

Example of failing checkpoint, taken from #40120 (needs num encoder layers != num decoder layers)

import torch
from transformers import AutoTokenizer, pipeline

torch_dtype = torch.float16
device_map = "cpu"
model_kwargs = dict(torch_dtype=torch_dtype, device_map=device_map)
model_id = "sshleifer/distilbart-cnn-12-6"
tokenizer = AutoTokenizer.from_pretrained(model_id)
generator = pipeline(
    "summarization",
    model=model_id,
    tokenizer=tokenizer,
    model_kwargs=model_kwargs,
)
generation_config = generator.model.generation_config
generation_config.do_sample = True
generation_config.use_cache = True
generation_config.temperature = 1.0
generation_config.num_beams = 1
generation_config.max_new_tokens = 100
generation_config.min_new_tokens = 100
generation_config.top_p = 1.0
generation_config.cache_implementation="static"

prompt = "I like math"

output = generator(prompt , batch_size=1, generation_config=generation_config)  # the cache will have an incorrect number of layers on `main`, but it runs
output = generator(prompt , batch_size=1, generation_config=generation_config)  # crashes on `main`

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[gante]

@Cyrilvallez CI green now 👌

(possibly there are a few more edge cases, but I don't think it's worth going the extra mile to find them all)

[manueldeprada]

LGTM, only comment I have is that it would be clearer for me to name it cross_attention_cache instead of encoder_cache and same for self_aattention_cache and decoder_cache.

[gante]

@manueldeprada like this? (see latest changes)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: colqwen2, dia, qwen2_5_omni, t5gemma