Provide more detailed instructions on readme

[Willian7004]

It takes more time to find the "[original args]" on the orginal readme, I suggest to provide a full command to run the example propmt and note that the supplementary explanations on the orginal readme.

The default batch size and cache size is too large for 6g VRAM, I suggest to use batch size=1 and cache size=12288 for 6g VRAM.

Based on the suggestions, I suggest using the following command by default :

python src/yue/infer.py --stage1_use_exl2 --stage2_use_exl2 --stage2_cache_size 12288 --cuda_idx 0  --stage1_model YuE-s1-7B-anneal-en-cot-exl2 --stage2_model YuE-s2-1B-general-exl2  --genre_txt ./prompt_egs/genre.txt --lyrics_txt ./prompt_egs/lyrics.txt --run_n_segments 2 --stage2_batch_size 1 --output_dir ./output  --max_new_tokens 3000 --repetition_penalty 1.1

[allo-]

I think it would also be good to clearly describe what affects quality. Does the cache size only affect performance or also quality? In another issue it also seemed like a too small cache size is a pitfall that may lead to a silent output.

The readme also suggests to run with original arguments and doesn't mention that one needs to change the model paths (and download the models locally). Maybe the whole example in the new section of the README could be self contained, i.e., repeating the command line with the additional parameters and changed model names.

Also --stage2_batch_size is documented in the source as "non-exl2" what may be relevant for the readme. At least to me it was not clear if it is used and if it has advantages with exl2 or not.