Reduce peak memory usage in clean_segmentation and _clean_frame

[Copilot]

clean_segmentation() duplicated the entire array upfront, and _clean_frame() accumulated all component boolean masks simultaneously — together causing severe memory spikes on large (T, Z, Y, X) stacks.

Changes

clean_segmentation() — inplace parameter

• Adds inplace=False parameter. Set True to skip the full array copy and halve peak memory:

  # default: safe copy (existing behaviour)
  cleaned = clean_segmentation(seg)
  
  # memory-constrained: modify in place
  cleaned = clean_segmentation(seg, inplace=True)

• Removes leftover debug code that printed frame-13 / label-31 diagnostics on every call.

_clean_frame() — label-by-label processing

• Drops the component_info dict that kept all component masks in memory at once.
• Processes one label at a time: del label_mask immediately after ndimage.label, compute sizes via np.bincount (no per-component arrays), then create and apply each comp_mask individually before moving on.

Tests

• Adds test_inplace_false_does_not_modify_input, test_inplace_true_modifies_input, and test_inplace_true_produces_same_result_as_copy to TestCleanSegmentation.