Support per-layer cache configuration and attention-mask selection

[eladsegal]

What does this PR do?

Updates cache construction and attention-mask selection to support models whose layers use different cache settings.

Previously, DynamicCache and StaticCache derived one set of initialization arguments from the top-level config and reused it for every layer. They now initialize each cache layer from its corresponding layer config, preserving per-layer sliding_window, attention_chunk_size, and number_of_conv_states values.

Cache and mask helpers also inferred cache metadata from a fixed layer, which could select a linear-attention layer or an unpopulated KV layer instead of one representing the requested attention pattern. When no explicit layer_idx is provided, they now select a representative KV layer: dynamic caches prefer a populated matching layer, while static caches use the first matching KV layer in their fixed layout.

Who can review?

@ArthurZucker
@Cyrilvallez

(Re-opening of #47901)

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 33315074346:2
Result: failure | Jobs: 16 | Tests: 137,956 | Failures: 0 | Duration: 12h 6m