Add Apertus adapter tests and fix nested key translation

[abhinav-bellapu]

Part of #1302.

Summary

• Add download-free focused coverage for ApertusArchitectureAdapter config routing, registration, component paths, GQA weight conversions, key translation, rotary/eager test setup, and the XiELU meta-device compatibility patch.
• Make Hugging Face-to-TransformerLens key translation prefer nested component paths over their parent prefix.
• Verify Q/K/V/O reshapes numerically round-trip and use the correct query vs. KV head counts.

Why

Apertus was the remaining unclaimed architecture in #1302 without a dedicated supported-architecture test suite.

The new key-translation cases also exposed a shared bug: self_attn or mlp matched before nested paths such as self_attn.q_proj and mlp.up_proj. This emitted keys such as blocks.0.attn.q_proj.weight instead of the canonical blocks.0.attn.q.weight, so the result was not the inverse of the existing TL-to-HF mapping and did not line up with adapter weight-conversion keys. Checking the more specific nested path first fixes that while preserving parent-component fallback behavior.

Validation

• uv run --locked --python 3.11 pytest tests/unit/model_bridge/supported_architectures/test_apertus_adapter.py -q (31 passed)
• pytest tests/unit/model_bridge -q (2404 passed, 6 skipped, 9 xfailed)
• pytest tests/unit -m "not slow" --reruns 2 --reruns-delay 5 -q (3260 passed, 33 skipped, 42 deselected, 9 xfailed)
• mypy . (Success: no issues found in 308 source files)
• pycln --check --all . --exclude __init__.py
• isort --check-only .
• black --check .