Need fp8 support to use H100/H200 and halve GPU costs

[samueldashadrach]

H100/H200 have dedicated fp8 tensor cores with twice as many FLOPs as fp16. This will halve GPU cost. Developers will use whichever library has lower cost, assuming the cost reduction comes without the dev having to put additional effort.

Please consider prioritising fp8 support, your library risks going outdated otherwise.

[samueldashadrach]

Update: I'm doing embedding inference using best models on MTEB benchmark, to be more specific.