Bump up the default value of --n_gpu_layers

[aendk]

The current default is 99 (e.g. in llama-bench). Some models already have around half of this.

99 also suggests to some people that this parameter could be a percentage or similar.
It being set to 99 confused me for a bit. Since it is stored as int32, I would suggest setting it to std::numeric_limits<int32>::max.

This would avoid confusion now, and incorrect testing once models grow to around 99 layers in the future.