Fix specparam NaN handling and add flexible z-score normalization options

[AlejandroOrozcoValero]

1. Fix ValueError in specparam peak selection

Problem: When computing power spectra with small nperseg values, some EEG signals produce all-NaN peak power arrays. The np.nanargmax() call on these arrays raises ValueError: All-NaN slice encountered, causing pipeline failures.

Solution: Added validation to check if the peak power array contains all NaN values before calling nanargmax(). If the array is all-NaN, peak selection is safely skipped.

Impact: Prevents pipeline crashes when processing signals with invalid power spectra. Samples with all-NaN peaks have no peak selected, which is the correct behavior. Valid samples are unaffected.

Modified file: ncpi/Features.py (lines 1343-1348)

2. Add per-epoch z-scoring and last epoch exclusion options

New configuration options:

zscore_after_epoch - Apply z-score normalization after epoching instead of before. This enables per-epoch normalization where each epoch is independently scaled to mean=0 and std=1. Default: False.
exclude_last_epoch - Exclude the last epoch from each time series. Useful for discarding potentially incomplete epochs when signal length is not perfectly divisible by epoch length. Default: False.

Implementation details:

• Modified processing pipeline to support conditional z-score application based on the zscore_after_epoch flag
• Added _exclude_last_epoch() method with proper handling of grouped time series
• Updated pipeline to apply last epoch exclusion after epoching but before optional post-epoch normalization

Modified file: ncpi/EphysDatasetParser.py

• Config options: lines 183-184
• Pipeline logic: lines 237-260
• New method: lines 959-978