Add TensorStore to performance benchmarks

[alxmrs]

I'm really excited to see the great read throughput improvements from Icechunk on Zarr v3! I'm curious how read throughput performance compares to TensorStore.

Having a place for consistent comparison would be really fruitful for Xarray data loaders, like xbatcher (xarray-contrib/xbatcher#42) (or eventually neuralgcm/neuralgcm#97.)

xref: google/tensorstore#49

[alxmrs]

Adding useful pointer to a comment (that might otherwise get lost):
https://github.com/neuralgcm/neuralgcm/blob/ddf63000d635e21be7dc38f3242d9d829af40fd9/neuralgcm/reference_code/reader.py#L395

[rabernat]

Thanks for this Alex! While we're excited about the initial benchmarking results, performance is not our current focus. We're still mainly focused on correctness, feature implementation, and stabilization of the file formats. So we won't be prioritizing this comparison in the near term.

We welcome you and anyone else to do comparisons if you wish, bearing in mind that basically zero effort has gone into performance optimization in Icechunk, and thus there remain many low-hanging fruits in terms of optimization.