Add ModernBERT support to MLX token classification export

[maziyarpanahi]

Summary

• Add ModernBERT token-classification support to the MLX backend, including rotary attention, alternating global/local windows, gated MLP blocks, and the classification head.
• Wire ModernBERT configuration and Hugging Face weight-key dispatch for MLX-fp and MLX-8bit artifacts.
• Add synthetic offline coverage for architecture dispatch, PyTorch/MLX logit parity, and INT8 output shape preservation.

Validation

• tests/unit/mlx/test_mlx_architecture_coverage.py — 3 passed on the arm64 MLX environment.
• tests/unit/mlx/test_mlx_convert.py — 34 passed, 2 skipped.
• Ruff lint and format checks pass for all changed files.

Closes #2345

[maziyarpanahi]

The ModernBERT work for #2345 has been consolidated into #2455, which is current with master, adds both ModernBERT and Longformer, and includes real Metal parity plus FP/INT8 artifact coverage. PR #2455 now closes both #2340 and #2345, so this narrower duplicate is superseded.