Generate: handle cache_position update in generate

[gante]

What does this PR do?

Updates cache_position in generate, and makes it the primary source for the input position in the models that support them, llama and gemma (as opposed to relying on past_key_values.seen_tokens).

The PR also adds the following related changes:

1. StaticCache now supports get_seq_length(). This was drawn from Static Cache: no mandatory cache_positions input #29221, and is needed for .prepare_inputs_for_generation() retrocompatibility;
2. The seen_tokens attribute enters a deprecation cycle, as it is redundant with cache_positions (and doesn't work with compilation).

This PR is drawn from the diff in #29374, i.e. it is a requirement for generate compilation with fullgraph=True 🙌

👉 Llama, Gemma, and Cache slow tests ran, no new failures
👉 FWD compilation benchmarks ran, no throughput change

[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.

[ArthurZucker]

Alright, I think Llama is already testing this. Moving fast here

[ArthurZucker]

alright, we are deprecating this anyways

[ArthurZucker]

my single worry here is potential stride, adding a .contiguous() might be needed

[gante]

I've double-checked, it's always (1,) 🤗 (which makes sense, since it's a 1D tensor)

Its shape will indeed be different, at least between prefill and subsequent generation

[ArthurZucker]

We should also set the dtype of the cache positions to int32 wdyt?

[gante]

Our integers inputs (input_ids, attention_mask, ...) are all int64, I think we should keep a consistent type :p

[ArthurZucker]

we have correct long typing here!

[gante]

(see int64 comment above)

[gante]

(rebasing and reruning tests, just in case 🙃 )