Add OLMoE architecture adapter tests

[RecreationalMath]

Description

Adds a unit test suite for OlmoeArchitectureAdapter under tests/unit/model_bridge/supported_architectures/, following the existing adapter-test pattern (modeled on the qwen2 / qwen3_moe / mixtral suites). It needs no model downloads or real checkpoints: it uses tiny programmatic TransformerBridgeConfig objects, plus small synthetic tensors and a fake attention module for the behavioral tests, so it runs on CPU in seconds.

The suite (50 tests) covers:

• Adapter config defaults (RMSNorm, rotary, gated MoE MLP, final_rms=False, eager attention).
• Weight conversions: the four QKVO weights (OLMoE has no projection biases), with GQA-aware head counts and the no-n_key_value_heads fallback.
• Numerical round-trips: the rearrange conversions are actually run on synthetic HF-shaped weight tensors, asserting the split-head output shapes and lossless reversion.
• Component-mapping structure, bridge types, and HF module paths, including the Q/K-norm RMSNorm submodules and the MoE bridge with its gate router.
• Factory registration and dispatch via select_architecture_adapter.
• GQA forward hook shapes: a fake attention module (carrying OLMoE's pre-reshape Q/K norms) wired into the bridge confirms Q surfaces n_heads while K/V surface n_key_value_heads.
• setup_component_testing rotary-embedding wiring, eager-attention forcing, and robustness on a minimal HF model.
• prepare_model in-place-clamp patching: a no-op when clip_qkv is unset, when set it wraps attention forward so clip_qkv is disabled during the call and restored afterwards, and it tolerates a model without layers.
• Architecture guards against drift.

Contributes to #1302 (OLMoE checkbox).

Type of change

• New feature (non-breaking change which adds functionality)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[RecreationalMath]

The failing check is the Othello_GPT notebook at Cell 4, the import transformer_lens cell, which is the same cell flagged previously. The mechanism here is a bit different from the rate-limit flake noted there: nbval fails it with Unexpected output fields from running code: {'stderr'}. The cell's saved output is empty, but in the CI environment the import emits a warning on stderr, so nbval sees an extra output field and fails the cell.

Running the exact CI command locally (pytest --nbval-sanitize-with demos/doc_sanitize.cfg demos/Othello_GPT.ipynb) passes all 13 cells with the same package versions CI used, so the stderr only appears in the CI runner environment. This PR only adds a new file under tests/unit/, so it can't affect that cell.

[jlarson4]

Thanks for taking this on @RecreationalMath! Your tests look great, I will merge them shortly.

The Othello GPT notebook failure is a known issue, I reran the CI test and your code passed on a second run. Thanks for documenting the source of the problem, I'll see if we can't get that cleaned up.