Add neural networks model performance

[VanDavv]

This PR adds a model performance to the available models' table.

It was measured by running $ python3 depthai_demo.py -cnn <cnn_name>, without additional switches, to provide the same results as users would get when trying out these networks.

We can add something like Max FPS providing an example of how to run the demo script (or some custom one) to achieve the best performance

[VanDavv]

One idea @szabi-luxonis / @themarpe - if we would be able to provide the number of available shaves via the API, we could compile the models dynamically with the most optimal amount of shaves. Currently, we default to 4 shaves but we can customize it manually with --shaves param in the demo script

[themarpe]

I agree - in that case I'd most likely go 1 step further if clean OpenVINO integration was possible and allow to just specify IR instead. Then pipeline would compile into blob with max number of shaves available.

Although that depends on 2 important factors:

1. Possible to expose a subset of resource manager which determines number of available shaves by asking nodes on pipeline about their properties.
2. Possible to cleanly integrate multiple OpenVINO versions and their subsets (we could also just do openvino for latest otherwise, etc...).

[VanDavv]

I love this idea, to be able to just provide the IR and compile the blob automatically - it would make it a lot easier for someone already familiar with OpenVINO to make a transition to DepthAI.

One additional issue I see here though is where the compilation should happen and how to ensure we can compile the blobs. We can use the online converter, but this requires an internet connection. We can use the OpenVINO PyPi package, but it's not available on some OS-es.
But overall I think that releasing this as an additional feature, that may improve the experience for some of our users, would make sense.
Thoughts @Luxonis-Brandon?

[themarpe]

@VanDavv
In this case we'd be including the actual OpenVINO inside our library, so compilation would happen locally. No internet access required. Basically it'd be same as using OpenVINO, just that in our case that would happen opaquely in the background.

As mentioned, this depends on integration difficulties of OpenVINO itself and multiple versions of it.

[VanDavv]

I'll go ahead and merge this PR as it will be useful as an overall preview of the available networks and help with navigation around them.
@themarpe @szabi-luxonis @Luxonis-Brandon - feel free to still review this PR if you want, I'll address any change requests in a separate PR

[Luxonis-Brandon]

Sounds great thank you!