chore: Replace metadata dict data_loss_ratio with measure-based equivalents

[cbueth]

This PR replaces pre_processed_metadata['data_loss_ratio'] and ['data_loss_ratio_blinks'] (EyeLink-only, computed during ASC parsing) with format-independent equivalents:

• Total data loss: pymovements.measure.data_loss('pixel', sampling_rate=sr, unit='ratio'). Per-trial via group_by, simple mean across trials
• Blink data loss: gaze.measure_events_ratio('blink_eyelink', sampling_rate=sr). Per-trial via group_by, simple mean across trials
  The old metadata values remain in the overview alongside the new ones for comparison during transition.

Instead of just changing this, I first had both be calculated in parallel (1351619), to see if the results are the same, or differ:

Comparison data

Here compare the results of the metadata numbers of before and the measure numbers of after.

Session	meta_total	meas_total	(\Delta)	meta_blink	meas_blink	(\Delta)
[redacted]	0.04633	0.06233	0.01600	0.04632	0.05794	0.01161
[redacted]	0.03367	0.03807	0.00440	0.03367	0.03553	0.00186
[redacted]	0.02036	0.02329	0.00293	0.02035	0.02182	0.00146
[redacted]	0.05463	0.06996	0.01533	0.05463	0.06547	0.01084
[redacted]	0.01961	0.02206	0.00245	0.01961	0.02068	0.00107

The ~1-2% difference comes from two sources:

1. Simple mean vs. weighted aggregate: the new code uses .mean() (trials equally weighted), the old metadata uses Σ(expected − valid) / Σ(expected) (weighted by trial duration per recording block).
2. Trial columns ≠ recording blocks: the measure groups by ['trial', 'stimulus', 'page'] while metadata uses EyeLink START/STOP recording messages. Boundaries differ.
   See my comment Remove data_loss_ratio and data_loss_ratio_blinks from parsed metadata pymovements/pymovements#1584 (comment) for more details on this.

Open question

Should we use a weighted aggregate instead of simple mean? The current .mean() treats each trial equally. To replicate the old metadata-style weighted aggregate, we would need to weight by trial duration. Replacing

.select(pl.col('data_loss_ratio').mean())

with

# Sum per-trial loss counts, divide by total expected
.agg(dl_count).sum() / .agg(dl_expected).sum()

Please add opinions on this when reviewing! Is the information saved where expcted?

References

• Remove data_loss_ratio and data_loss_ratio_blinks from parsed metadata pymovements/pymovements#1584
• Store calibration and other metadata in well-documented form pymovements/pymovements#893
• improve null_ratio functionality and deprecate parsing._calculate_data_loss pymovements/pymovements#1082
• implement data loss ratio as a sample measure pymovements/pymovements#1398
• feat: add sample measure data_loss with count, time, and ratio units pymovements/pymovements#1420

[theDebbister]

Ok, so in consultation with the other MultiplEYE team mebers we decided that it is best to report the session level measures (in general) not as simple averages over the trials, but taking the length into account (i.e. just calculate the data loss at the session level directly).

As the other values can simply be calcualted based on values we already have at the trial level, we can later still add these, but for now, we just use this one value.