Welcome Tobias and thanks for posting an interesting question! I’m sorry I didn’t reply earlier…COVID got me last week. 
I strongly favor the direct Zarr 
IPLD approach, bypassing IPFS. As you note, it is simpler and cleaner, avoiding the unnecessary intermediate abstraction of files and blocks.
As we have learned recently thanks to the work by @martindurant, @lsterzinger, and @rsignell with kerchunk, basically every common legacy file format (netcdf3, netcdf4, grib, geotiff / COG) can be mapped 1:1 to the Zarr data model. If we had a way to way to store the Zarr data model in IPLD, we could effectively bring all of these formats along for the ride.
I’m curious: what metrics could we use to help decide among the two approaches? Performance? Convenience? Support in the ecosystem?