Add anonymization, Studio, and Nemotron privacy-filter support

[maziyarpanahi]

Summary

• Adds a Faker-backed anonymizer with locale-aware, deterministic, and format-preserving surrogate generation.
• Introduces canonical PII label normalization plus shared BIOES decoding and span refinement utilities for privacy-filter backends.
• Adds cross-platform privacy-filter routing with MLX support on Apple Silicon and PyTorch fallback elsewhere.
• Extends the public privacy-filter family with Nemotron-PII checkpoint IDs: OpenMed/privacy-filter-nemotron, OpenMed/privacy-filter-nemotron-mlx, and OpenMed/privacy-filter-nemotron-mlx-8bit.
• Adds family-aware Torch fallback so Nemotron MLX requests on non-MLX hosts substitute OpenMed/privacy-filter-nemotron instead of openai/privacy-filter.
• Adds classifier-head bias support for Nemotron MLX artifacts while preserving the original bias-less OpenAI Privacy Filter default, including native Swift OpenMedKit artifact decoding/model construction.
• Switches the OpenMed Scan Demo privacy-filter engine to OpenMed/privacy-filter-nemotron-mlx-8bit and updates the visible engine naming/docs.
• Adds Privacy Filter Studio, an interactive FastAPI/static web demo for masking or deterministic Faker randomization with sample notes, highlighted entities, backend/model status, and explicit first-run download control.
• Adds Portuguese API schema support, trust-remote-code loading for the Torch privacy-filter family, docs, demos, changelog/readme updates, lockfile refresh, and focused unit coverage.

Review Notes

• Reviewed the new Swift OpenMedKit classifier-bias decoding/model changes, Scan Demo Nemotron switch, Studio app/config, MLX classifier-head change, changelog entries, routing, MLX dispatch aliases, Torch loader changes, docs, example, and tests before committing.
• Fixed Studio behavior before commit so cache-only mode actively sets Hugging Face offline flags and model-load/inference errors render as errors in the output pane.
• Earlier docs fixes in this PR covered artifact count wording, tiktoken typo, baseline-vs-fine-tune wording, and fuzzy changelog validation count.

Validation

• git diff --check
• swift test --filter OpenMedMLXTests/testPrivacyFilterConfigDecodesNemotronClassifierBias (1 passed)
• swift test --filter OpenMedMLXTests/testPrivacyFilterConfigAcceptsUnembeddingBiasAlias (1 passed)
• swift test --filter OpenMedMLXTests/testTinyPrivacyFilterNemotronModelForwardShape (1 skipped: MLX runtime resources unavailable from SwiftPM CLI test bundle)
• swift test --filter OpenMedMLXTests/testTinyPrivacyFilterBaselineHasNoUnembeddingBiasSlot (1 skipped: MLX runtime resources unavailable from SwiftPM CLI test bundle)
• .venv/bin/python -m compileall -q examples/privacy_filter_studio openmed/mlx/models/privacy_filter.py
• FastAPI Studio smoke via TestClient: GET /api/examples -> 200, empty POST /api/run -> 200 / empty
• .venv/bin/python -m pytest tests/unit/mlx/test_privacy_filter_mlx.py tests/unit/test_privacy_filter_routing.py (20 passed, 8 skipped)
• Earlier focused privacy/anonymization suite: .venv/bin/python -m pytest tests/unit/core/test_anonymizer.py tests/unit/core/test_labels.py tests/unit/test_pii.py tests/unit/test_privacy_filter_routing.py tests/unit/test_pii_multilingual_regression.py tests/unit/mlx/test_privacy_filter_mlx.py tests/unit/service/test_api.py (471 passed, 1 skipped, 11 warnings)