Session reliability metrics for cue and lever epochs #173

Add trial-to-trial reliability metrics to the _metrics-session.csv written by process metrics, for go/no-go sessions only.

Epochs

• Cue-aligned (cue_onset, trial_start): each trial is masked from its response time on. Miss and correct-rejection trials end at the median RT of all responding trials. --no-mask-response turns masking off.
• Lever-aligned (response): the epoch runs from the cue to the lever push, end exclusive, so the movement and reward are left out. Needs process peri-event --event response aligns to start_time + response_time instead of cue_onset + response_time #176.
• Trials with RT < 200 ms are dropped (--min-rt).
• Lever-aligned baseline is taken before each trial's cue. Trials whose pre-cue baseline falls outside the window get NaN, and the count is reported. Lever-aligned runs need a longer --pre.

Metrics (new columns; existing columns unchanged)

• epoch_correlation: zero-lag mean pairwise Pearson r over the masked epoch, using the samples each pair shares. --time-warp resamples each epoch onto a common 0–1 grid instead.
• response_fraction: fraction of trials whose mean epoch ∆F exceeds baseline by --response-sd (default 2) baseline SDs.
• variance_quench: across-trial variance over the epoch divided by across-trial variance over the baseline.
• signal_fraction: fraction of single-trial variance explained by the trial mean.
• reliability_n: trials used after exclusions.

Groups (wide format, still one row per region)

• <metric> covers all trials.
• <metric>_sdt-<hit|miss|false_alarm|correct_rejection> gives one column per trial type.
• <metric>_stim-<go|nogo> gives one column per stimulus type.
• Every group column is always written. A group gets NaN when it has fewer than 10 trials or no trials.
• If a peri-event CSV lacks sdt_type or response_time, the new columns are skipped with a warning.

The peri-event report's session table shows the new columns, with a group selector. Documented in docs/how-to/metrics.md.