feat: allow to preload ML models when running inference

[Angelmmiguel]

Introduce the ML model preloading by leveraging the "Named models" feature from WASI-NN. Now, workers don't need to load the model files and send it to the host, but loading it in the host directly. I added a new example that uses this feature and produces the same result as the rust-wasi-nn example.

I also introduced some new configuration properties to the wasi_nn feature:

name = "wasi-nn-mobilenet"
version = "1"

[[features.wasi_nn.preload_models]]
backend = "openvino"
provider = { type = "local", dir = "./_models/mobilenet/" }

The preload_models array allows to configure multiple models. The provider configures the way wws retrieves the models. Currently, only the local provider is available but we plan to add more.

Limitations

Even though we're preloading the models on the host, this process still happens at request level. Based on some comments, the Wasmtime WASI-NN library is not protected against concurrent access. To avoid multiple workers accessing the same context, I'll keep this at request level for now.

In the future, my goal is to load models once at worker initialization.

It closes #215

[ereslibre]

Super!, LGTM :)

[ereslibre]

Nice :)