Bug: EvalConfig preset resolver aliases shared list, mutation corrupts BENCHMARK_PRESETS

[sacredvoid]

Description

In `eval.py:51-53`, the preset resolver assigns the `BENCHMARK_PRESETS` list directly:

```python
self.tasks = BENCHMARK_PRESETS[self.preset] # aliasing, not copying!
```

Since both point to the same list object, any mutation of `config.tasks` (e.g. `config.tasks.append("extra")`) permanently corrupts the shared preset dict for all future configs in the same process.

Reproduction

```python
from alignrl.eval import BENCHMARK_PRESETS, EvalConfig
cfg = EvalConfig(preset="reasoning")
cfg.tasks.append("extra")
print(BENCHMARK_PRESETS["reasoning"]) # ['gsm8k', 'math', 'arc_challenge', 'extra'] - CORRUPTED
```

Impact

In notebooks where multiple EvalConfig instances are created (e.g. evaluating different stages), modifying one config's tasks silently corrupts all subsequent preset resolutions.