[BUG] ngmix crash

[martinkilbinger]

@fabianhervaspeters has reported seeing crashs of ngmix after a batch of 1000 object is written to file.

We should check that batch writing and reading is well implemented.

[cailmdaley]

Root cause found and reproduced. The crash is a column-dtype mismatch at the batch-append boundary, not a write/read logic error:

• compile_results fills nfev_fit with np.nan when a fit type failed — the only int-like output column with a float fallback (every other NaN-filled column is float from batch 1).
• With SAVE_BATCH enabled, save_results locks all column dtypes from the first batch. If batch 1 contains no failed fits, nfev_fit locks to int64; the first later batch containing a failure then assigns a NaN into that int64 structured column and raises ValueError: cannot convert float NaN to integer — always right after a batch is written, matching @fabianhervaspeters's observation.

Fix in #802: fall back to ngmix's own -1 failed/absent-nfev sentinel instead of NaN, keeping the column int64 across every batch. Comes with a regression test that drives the real two-batch save_results path — it fails on current develop with the exact reported error and passes with the fix. Module + fast-science suites green.

Two notes: the bug affects any SAVE_BATCH run (not image-sims-specific), and whether a run crashes depends only on whether its first batch happened to be failure-free — which is why it appears intermittent across tiles.

— Claude Fable on behalf of Cail