fix: set-list zip produces incorrect span pairings in PerClassScorer (candidate_generation.py)

## Problem

In `scispacy/per_class_scorer.py` line 21-23, `untyped_predicted_spans` is constructed as a **set**:

```python
untyped_predicted_spans = {(x[0], x[1]) for x in predicted_spans}
```

Then on line 23, it is **zipped** with the original list:

```python
for untyped_span, span in zip(untyped_predicted_spans, predicted_spans):
```

**Sets do not preserve insertion order in Python.** This means the `untyped_span` variable may not correspond to the correct `span` from `predicted_spans`. The result is that the untyped TP/FP counting (lines 30-34) can attribute matches to the wrong spans, producing incorrect untyped precision/recall metrics.

## Proposed Fix

Change the set comprehension to a list comprehension (or compute the untyped span inline):

```python
untyped_predicted_spans = [(x[0], x[1]) for x in predicted_spans]
```

## Affected File

`scispacy/per_class_scorer.py`