Add drift QC: per-FOV drift_{fov}.csv and fishtank plot-drift

[leonardosepulveda]

Summary

Adds drift quality-control output and a plotting command to detect-spots.

• Per-FOV drift QC from channels_{fov}.csv. detect-spots already records the
  per-series integer-pixel drift (x_drift, y_drift) it applies during intensity
  quantification in channels_{fov}.csv. This PR reuses that record for QC rather than
  recomputing or duplicating it.
• fishtank plot-drift. A new CLI subcommand (and fishtank.pl.plot_drift function)
  that reads the per-FOV channels_*.csv, de-duplicates to one drift value per
  (fov, series), and renders a three-panel QC figure:
  1. x vs y drift scatter, one point per (fov, round), colored by round, on square
     symmetric axes (with an outlier-resistant --scatter_max auto option);
  2. histogram of drift distance (computed from x_drift/y_drift) with the median marked;
  3. a fov x round heatmap of drift distance − median (diverging colormap).

Rationale for revisions

The original version of this PR measured drift at sub-pixel resolution
(phase_cross_correlation(..., upsample_factor=10)) and wrote a separate
drift_{fov}.csv table. Both were reverted:

• Sub-pixel registration was dropped. upsample_factor=10 significantly slows down the
  drift calculation, and since the drift applied to images stays integer-pixel anyway, the
  extra precision was never used — it only fed the QC record. detect-spots is back to the
  original integer-pixel phase_cross_correlation(...).
• drift_{fov}.csv was removed as redundant. The fov, series, x_drift, and
  y_drift columns it contained are already written to channels_{fov}.csv. plot-drift
  now reads channels_{fov}.csv directly (deduping to one row per (fov, series)), so there
  is a single source of truth and no extra output file.

Notes

• Non-breaking: no new output files or columns; detect-spots behavior is unchanged.
• Plotting uses only existing dependencies (matplotlib, numpy, pandas).
• The figure defaults to <input>/drift_qc.png; axis limits, histogram bin width, colormap
  and font size are all parameters (with auto limits available for scatter/histogram).

🤖 Generated with Claude Code

[colganwi]

@leonardosepulveda I removed the drift_{fov}.csv file since its redundant with channels_{fov}.csv and removed the upsampling since its significantly slower. Let me know if there was a rational for upsampling beyond detecting 0 drift which should have been solved by your other PR.