The random walk random field version deviates quite a bit from the IID as.factor(year) model and I can't see an obvious reason. The IID version makes more sense to me eyeballing the raw data, and I think I would be inclined to go that direction here since there is no obvious deviation from the survey plan or combining of a patchwork of surveys. Their is most definitely an increase in zeros towards the end in the raw data, which is reflected in the IID version. The survey timing is potentially problematic, but including day of year actually doesn't do much. If you look at the raw data you'll notice that early surveys are associated with both high (1993) and low (2003) years. A plot of index vs. day of year doesn't show much.
Adding notes here: