make_described_run writes image_count images per technical repeat instead of per member #234

This blocker comes from the external review of #228 (ChatGPT batch 32, claim A3).

testing_described.py:89-95 yields member.image_count once for each technical repeat, and :128-152 writes that many images into each unit. But image_count belongs to the member: it is the total over the member's units, and reconciliation sums the unit counts and compares the sum with it (description_reconcile.py:242-263).

In fixture v34, CTRL/rep1 has image_count = 12 over two technical repeats, so the helper writes 24 images instead of 12. DrugA/rep1 has 11 and gets 22. The regression (tests/test_make_described_run.py:25-32) computes its expected count the same wrong way, so it passes while the biological N is wrong.

Fix:

• Spread exactly member.image_count images across the member's technical repeats, deterministically.
• Every technical repeat gets at least one image. If image_count < len(technical_repeats), refuse to build that fixture rather than invent an empty unit.

Regression: build v34 and check against counts computed independently of the helper:

• CTRL/rep1 has exactly 12 FOVs over acq1 and acq2.
• DrugA/rep1 has exactly 11.
• Both acq1 and acq2 appear.
• For every (condition_id, repeat_group), the total FOV count equals the parsed member.image_count.