perplexity / imatrix / kl-div is not being calculated properly on Windows without a special flag

[kalomaze]

Encountered a pretty nasty bug that helps explain some things about my past efforts w.r.t calibration datasets.

perplexity.exe -m "C:\Users\Kalo\Documents\GitHub\llamacpp_git\examples\dpo_7b_quant\DPO_7b_q8_0.gguf" -f 8k_redo.txt -ngl 33 -c 512

On Windows, supplying -f to the perplexity example (and presumably imatrix as well which is derivative of it) results in higher perplexity because it is not reading it properly as a binary file like how it does on WSL / Ubuntu (regardless if you use -f or -bf, it is read properly there).

Here is about 9k tokens worth with -f on Windows:

perplexity: calculating perplexity over 18 chunks, batch_size=512
perplexity: 0.40 seconds per pass - ETA 0.12 minutes
[1]8.9604,[2]13.1455,[3]12.9100,[4]12.6800,[5]13.6468,[6]14.1241,[7]14.4815,[8]13.3205,[9]12.5099,[10]12.2798,[11]11.8491,[12]11.4355,[13]11.3987,[14]11.4812,[15]11.5342,[16]11.6241,[17]11.2636,[18]11.1466,
Final estimate: PPL = 11.1466 +/- 0.46029

And here is -bf for comparison:

perplexity: calculating perplexity over 19 chunks, batch_size=512
perplexity: 0.40 seconds per pass - ETA 0.12 minutes
[1]8.1417,[2]11.9650,[3]12.0453,[4]11.8829,[5]12.6807,[6]11.1073,[7]11.8471,[8]11.5785,[9]10.6265,[10]11.1294,[11]10.6127,[12]10.1494,[13]10.0213,[14]10.1782,[15]10.2959,[16]10.5644,[17]10.4623,[18]10.3628,[19]10.3186,
Final estimate: PPL = 10.3186 +/- 0.42002

In addition to this, KL divergence was not working properly on Windows whatsoever, because the output file wasn't being read as a binary file. I have made a PR to address this:
#5273

This also makes me concerned that imatrix calculations are not working as intended either on Windows; maybe even if you use -bf. I will be investigating that shortly