Fix probe attachment in serialization fixture builders

[h-mayorquin]

I got this error on #4840 (Filter one channel at a time to reduce peak memory) and decided to fix it here. When we changed set_probe() and set_probegroup() to modify the recording in place, calls without in_place stopped returning a recording. The fixture builders account for removing the argument but still assign the return value to rec_with_probe. That value is None, so the script fails when it tries to save the recording.

This was working before because the script uses in_place=False for versions up to and including 0.105.0. That release still supports the deprecated argument and returns a recording in that case. The workflow automatically selects the latest published patch of each minor version. Once 0.105.1 was published, it selected the other path in the script and exposed the incorrect return-value assumption. The in-place behavior was already present in 0.105.0. It was the version selected by the workflow that changed.

This PR keeps the original recording after attaching the probe in the newer-version paths. The older-version paths are unchanged. I checked fixture generation and all 26 round trips separately with 0.105.0 and 0.105.1, including parallel binary and Zarr writes.