Eliminate toarray() for spectral/spatial metrics

[Bisho2122]

Problem

Follow-up to #1718 . The annotation pipeline sometimes crashes with OOM on datasets where spectra form isolated blobs inside a large bounding box. After #1718 fixed chaos_metric, the remaining crash point is formula_validator.py which still materialises the full h×w dense array for every peak of every formula when building iso_imgs_flat.

Change

• imzml_reader gets a pixel_to_flat_idx lookup table built once at init — maps any pixel coordinate directly to its position in the
  masked-flat metrics array.
• formula_validator.py uses it to scatter sparse coo values directly into a 1D array of size n_spectra, removing the toarray() call and the intermediate dense iso_imgs list entirely.
• pipeline.py logs a warning when pixel density is below 5% so blob datasets are visible in logs.