Hotfix for relevance model

[p-ferreira]

It seems that we need self.model.cuda() to bring the model into memory. This function brings the model to the device specified on the model initialization, and it's defined in the angle lib as:

    def cuda(self):
        if self.load_kbit is None:
            self.backbone = self.backbone.to(torch.device(self.device))
        return self

Without model.cuda(), the test validator was throwing the classic exception tensors are not in the same device (gpu and cpu) on encoding

Other changes included in the PR:

• normalizes configs for miners
• adds README for wiki agent and zephyr miner
• overall adjustments on dependencies and wiki agent

[steffencruz]

Amazing. Looks good to me