v0.2.0

Routes each request to the checkpoint best suited to it.

## New

**`Router`** picks between `laya` (English), `laya-multilingual` (100+ languages) and `laya-typed-decisions`, loading lazily with LRU eviction.

```python
from laya import Router
router = Router()
router.predict({"body": "I was charged twice"}, questions)          # -> laya
router.predict({"body": "मुझसे दो बार शुल्क लिया गया"}, questions)   # -> laya-multilingual
router.predict(state, questions, model="typed-decisions")           # explicit
```

`route()` returns the decision and its reasoning without loading anything.

**`laya.lang`** — exact Unicode script detection over 22 scripts plus a best-effort Latin language guess. No new dependencies.

## Fixed

- **`laya-multilingual` could not be loaded at all.** Its `extra_special_tokens` ships as a list where transformers expects a mapping, raising `AttributeError: 'list' object has no attribute 'keys'`.
- ModernBERT's `reference_compile` now disabled — it torch.compiles the encoder at batch sizes where that is a loss, and can hang.
- Dropped a duplicate `email_questions` import that shadowed the presets one.

## Why route

Measured on a T4, identical questions per model:

| | `laya` | `laya-multilingual` |
|---|---|---|
| MASSIVE intent, English | **0.783** | 0.657 |
| MASSIVE intent, 13 others | 0.306 | **0.451** |
| XNLI, 14 non-English | 0.521 | **0.731** |

Across all 51 MASSIVE languages the English checkpoint macro-averages 0.227 and clears 3× random on only 23 of 51; the multilingual checkpoint reaches 0.366 and clears it on 45. Khmer scores 0.000 accuracy at 0.952 confidence — the model's own confidence gives no warning, so the decision has to happen before the forward pass.

## Note

If `laya.load()` hangs and TensorFlow is installed, run with `USE_TF=0`. transformers probes for TF at import and its abseil runtime can deadlock model construction.

**Full benchmark data:** the `research` branch.