Support for NIR <> lava-dl

[stevenabreu7]

User story

As a user, I want to be able to train SNNs in other libraries (e.g. snnTorch, Norse, Nengo, Rockpool, Sinabs), export the trained networks and import them in lava-dl to then run them on Loihi 2. The neuromorphic intermediate representation (NIR) project aims to make neuromorphic simulators and hardware platforms more interoperable. I would like to add support for converting between lava-dl and NIR (bi-directional).

Conditions of satisfaction

• Users should be able to convert NIR graphs (generated from any other neuromorphic simulator) to lava-dl networks.
• Users should be able to take a network trained in lava-dl and export it to NIR.
• At first, supporting the set of networks that are supported by both NIR and lava-dl suffices. Later on, we may extend NIR to support other neuron models (e.g. sigma-delta).

Acceptance tests

• Converting lava-dl → NIR → lava-dl should not lose information or break the lava-dl network.
• Similarly, converting NIR → lava-dl → NIR should not lose information or break the NIR network.
• Demonstrate that conversion between lava-dl <> NIR works for the lava-dl example networks (pilotnet-snn and oxford).