Publish detector view images as float32

[SimonHeybrock]

Fixes #1290.

da00 is uncompressed, so the image dtype is the wire size. Detector views published float64 because ToNXevent_data allocates its per-event weight buffer as float64 and bins.sum() carries that dtype through to the image — every event contributing a weight of exactly 1.0 in eight bytes. Halving the payload also halves what the dashboard's buffers hold per frame (#1274).

Measured through the real da00 path: a 640x640 panel drops from 3.29 MB to 1.65 MB per update, so NMX's three panels go from 9.9 MB to 4.9 MB. A float32 array roundtrips through scipp_to_da00/da00_to_scipp identically.

The cast is at the publishing boundary rather than at the source. Making the weights float32 would make the accumulated histogram float32, and float32 stops incrementing at 2**24 — a long-running cumulative image can reach that and silently stop counting. Casting the result instead rounds a displayed value by at most one part in 16 million, with nothing accumulating on top. So the histogram, both accumulators, the count scalars, spectrum_view and roi_spectra all stay float64; only the two count images are cast.

Scoped to the count images rather than Da00Serializer, since other da00 payloads carry physical quantities and timestamps where float32 is not automatically appropriate.

The float64 per-event weights buffer is left alone: #1290 scopes it out as a separate saving, and flipping it is precisely what would make the accumulation float32 — it would need the accumulator to upcast deliberately.

Test plan

New unit tests cover the image being float32 with and without pixel weighting (the weighted path matters: weights are already float32, so image / weights promotes back to float64 before the cast), and the accumulated histogram staying float64 and still counting exactly past 2**24. The factory-level integration test additionally pins the count scalars as not cast.

Four existing tests compared an image against float64 via sc.isclose/sc.full, which raises DTypeError on mixed dtypes rather than failing a comparison; they now compare values or the expected dtype.

Unrelated to this change, five bifrost qmap tests fail identically on the unmodified branch point in a freshly resolved venv — ess.spectroscopy's monitor_wavelength_data is missing a keep_event_time_offset argument.

Worth a look in a running dashboard, since this changes what every detector view plot receives:

• Detector view images still render with correct values and color scaling.
• ROI selection and its spectra still behave (ROI spectra are unchanged, but they are computed from the same histogram).

[SimonHeybrock]

LGTM