feat: sort cuda graphs in descending order

[drbh]

sorting cuda graphs in descending order uses slightly less memory during initialization

using the following command on 1 A10G

text-generation-launcher \
--model-id meta-llama/Llama-2-7b-chat-hf \
--num-shard 1 \
--cuda-graphs 2,4,8,16,32,64,128,256,512,1024,2048

without cuda graphs sorted 22140/23028 MiB are used and with them sorted 21938/23028 MiB are used.

This is a small percent of total memory (<1%) however it's 202 MiB saved. Additionally looking at the memory usage overtime, loading in descending order has smaller spike (never goes above its final value, where in ascending it peaks higher than 22222 MiB)

Below are the two memory recording with https://github.com/drbh/nvline, left is ascending, and right is descending.

[drbh]

@danieldk thanks for suggesting this and reviewing!