Ollama ps command for showing currently loaded models

[pdevine]

This change adds a rudimentary ps command which makes use of the new scheduler changes in the server. The UX also

The UX for this depends on whether you're using the CPU, GPU, or a hybrid of both and looks like:

NAME            ID              SIZE    PROCESSOR        UNTIL
mistral:latest  61e88e884507    5.4 GB  100% GPU         28 seconds from now

NAME            ID              SIZE    PROCESSOR        UNTIL
mistral:latest  61e88e884507    5.4 GB  48%/52% CPU/GPU  28 seconds from now

NAME            ID              SIZE    PROCESSOR        UNTIL
mistral:latest  61e88e884507    5.4 GB  100% CPU         28 seconds from now

Additionally, there is a new --keepalive flag in the REPL which can be used to set how long you want the model to stay resident in memory after the model has finished inference. It takes a duration string (e.g. 3m30s), however we can switch this to also accept integers similar to the API.

This also introduces a new /api/ps endpoint which returns back a response similar to the /api/tags endpoint albeit with additional information. The size of the running model will not match the amount reported from the /api/tags endpoint for a given model since it can take additional memory when loaded onto the GPU or as a hybrid.

Partially addresses #3902
Fixes #4013
Replaces #2359

[jmorganca]

Suggested change

	keepAlive, err := cmd.Flags().GetString("keepalive")
	keepAlive, err := cmd.Flags().GetString("keep-alive")

I think this is ok as is, but would suggest keep-alive to be consistent with the api keep_alive – although the design of ollama's cli should be a matter of taste/usability over consistency

[pdevine]

I thought about this, but we have --nowordwrap as one of the other options for run, so it felt weird going to kebab case here.