Upgrade arrow2/convert and use native buffers for the tensor u8 types

[jleibs]

Depends on upstream: DataEngineeringLabs/arrow2-convert#103

Using Buffer types here should keep us from making a copy on deserialize.

TODO:

• Show this works based on memory profiling.

Checklist

• I have read and agree to Contributor Guide and the Code of Conduct
• I've included a screenshot or gif (if applicable)
• I've added a line to CHANGELOG.md (if this is a big enough change to warrant it)

[emilk]

Converted to draft until we have an arrow2_convert release

[emilk]

Sadly, this does not help with the memory duplication of tensors.
This can be seen with RERUN_TRACK_ALLOCATIONS=1 cargo r -p test_image_memory:

We see that the archive step still creates a complete copy of the memory.

The only real solution to this is:

• Save .rrd from store instead of saving LogMsg:es. #1394

[Wumpf]

@jleibs do we have the necessary arrow2convert version by now?

[emilk]

No; there hasn't been a new arrow2-convert for over two months. It is time we fork arrow2-convert and make our own version of it.