[Dataset Request] Add AI4Privacy pii-masking-300k — PII detection/masking benchmark

Hi! I'd like to request indexing the **AI4Privacy pii-masking-300k** dataset and a leaderboard for **PII Detection / Masking**.

### Dataset
- **HF:** [`ai4privacy/pii-masking-300k`](https://huggingface.co/datasets/ai4privacy/pii-masking-300k)
- **Size:** ~225K+ examples (multilingual; English subset commonly used)
- **Labels:** 27+ PII types (names, emails, phones, addresses, etc.); token-level spans + masked text
- **Task:** PII Detection / Named-entity–style span extraction
- **Primary metric:** token/span-level F1

### Suggested results to seed the leaderboard
Verified from the model cards. ⚠️ **These are not directly comparable** — they were trained/evaluated on different versions of the ai4privacy family and report different metrics. Please seed them as separate entries keyed to the correct dataset version:

| Model | Metric | Value | Dataset version | License | Source |
|---|---|---|---|---|---|
| `Isotonic/deberta-v3-base_finetuned_ai4privacy_v2` | overall token F1 | **97.57%** (P 97.22 / R 97.92) | pii-masking-**200k** | CC-BY-NC-4.0 | [card](https://huggingface.co/Isotonic/deberta-v3-base_finetuned_ai4privacy_v2) |
| `iiiorg/piiranha-v1-detect-personal-information` (mDeBERTa) | multiclass F1 | **93.12%** (headline metric on card is 99.44% *accuracy*) | pii-masking-**400k** | CC-BY-NC-ND-4.0 | [card](https://huggingface.co/iiiorg/piiranha-v1-detect-personal-information) |

Note both are self-reported model-card numbers, not from an independent leaderboard. More can be imported from `pwc-archive/evaluation-tables` if present.

### Note
One of the most widely used open PII benchmarks; a leaderboard here would be valuable for the privacy/NER community. It may be worth structuring per dataset version (200k / 300k / 400k) since models target different releases. Thanks!

The issue names no repository files, tests, or entry points. Start by finding how datasets and leaderboard results are indexed, then verify the pii-masking-300k metadata and determine how the 200k, 300k, and 400k versions should remain separate. Done means the requested dataset and leaderboard structure are represented without treating the self-reported model-card metrics as directly comparable.

Written by the indexing model from the issue text.