Fix NemotronH: Register "mlp" in the cache layer-type mappings

[qgallouedec]

NemotronHConfig emits four layer types (linear_attention, moe, full_attention, mlp), but cache_utils.py only knows three ("moe" is registered as a cache-less placeholder) "mlp" was left out.

So building any cache for a Nemotron-H config with - (MLP) blocks raises KeyError: 'mlp'.

That breaks the whole original Nemotron-H family (nvidia/Nemotron-H-8B-Base-8K, nvidia/Nemotron-H-4B-Instruct-128K, pattern M-M-M-M*-...) — generate() fails before any forward, as well as NemotronHConfig()'s default layer_types.

Fix: register "mlp" next to "moe" in DYNAMIC_LAYER_TYPE_MAPPING and STATIC_LAYER_TYPE_MAPPING. From the cache's point of view they are the same thing: a feed-forward block with no per-token KV and no recurrent state.

Regression from #47118.

Closes #47534

[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.

[Cyrilvallez]

Indeed, thanks a lot!

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 30253557001:1
Result: success | Jobs: 16 | Tests: 175,157 | Failures: 0 | Duration: 14h 3m