perplexity: update README FP16 results [no ci]

[JohannesGaessler]

The logits used for comparative runs of perplexity are stored as uint16_t instead of float. The difference from this downcasting can be non-negligible when looking at quants like q8_0 or q6_K. This PR adds a disclaimer and results to estimate the impact of the downcasting.

[fedric95]

@JohannesGaessler great work, plans to add also the scoreboard for llama3-70b? It would be very useful to compare the trend in perplexity loss of llama2-70b and llama-3-70b

[JohannesGaessler]

I'm hesitant to publish anything with LLaMA 3 70b because it turned out that the machine I built with 6x RTX 4090 has stability issues which means I have to be very careful that the data isn't being affected by random bit flips.

[fedric95]

I'm hesitant to publish anything with LLaMA 3 70b because it turned out that the machine I built with 6x RTX 4090 has stability issues which means I have to be very careful that the data isn't being affected by random bit flips.

:-( how much time does it take to run all the experiments for llama3 70b (more or less)? Just to understand how much it would cost.

[JohannesGaessler]

At standard settings a single LLaMA 3 70b run takes ~6 minutes on 6x RTX 4090.