AI: `-nvidia "all"` is not obvious what it does

Nobody has claimed this yet.

**Describe the bug**
When the flag `-nvidia "all"` is used for AI inference, the `aiModels.json` must match number of GPUs installed, else it will try run all GPUs with the first item in the list.

Not sure if this is the expected behavior.

**To Reproduce**
1. Install 2 GPUs in a machine
2. Set `-nvidia "all"`
3. Set your `aiModels.json` to include only the LLM model
```
[
{
"pipeline": "llm",
"model_id": "meta-llama/Meta-Llama-3.1-8B-Instruct",
"price_per_unit": 80000000,
"pixels_per_unit": 1000000,
"warm": true
}
]
```
4. If the second GPU has less than 24GB, it will fail to launch to container and time out.

**Expected behavior**
Obviously if we specify the GPUs ie `-nvidia 0,1` then we assume the `aiModels.json` will have two models loaded, it works.
i.e.
```
[
{
"pipeline": "llm",
"model_id": "meta-llama/Meta-Llama-3.1-8B-Instruct",
"price_per_unit": 80000000,
"pixels_per_unit": 1000000,
"warm": true
},
{
"pipeline": "text-to-image",
"model_id": "ByteDance/SDXL-Lightning",
"price_per_unit": 4768371,
"warm": true
}
]
```
I think the documentation needs to be updated, but I am unsure if this is actually the logic for the `"all"` value.

**Set Up**
Slot 0 - 3090 (24GB Ram)
Slot 1 - 2080 ti (11GB Ram)

Start by reviewing the documented behavior of the `-nvidia` flag, especially the `"all"` value, alongside the `aiModels.json` example in the repository. Confirm whether the model list must match the number of GPUs and document the behavior and constraints clearly; the issue provides no specific documentation file or test to run.

Written by the indexing model from the issue text.