Error Retrieval calculation error & Fix some detailed problems

[ztxtech]

Change outdated methods and fix some modules that default to specifying CUDA as the device.

[salesforce-cla]

Thanks for the contribution! Before we can merge this, we need @ztxtech to sign the Salesforce Inc. Contributor License Agreement.

[ztxtech]

There is a calculation error here. The Attention value should be weighted and memorized instead of the corresponding index.

Original code at fsnet_.py (line 105-108)

...
            v, idx = torch.topk(att, 2)
            ww = torch.index_select(self.W, 1, idx)
            idx = idx.unsqueeze(1).float()
            old_w = ww @ idx
...

Corrected code

...
            v, idx = torch.topk(att, 2)
            ww = torch.index_select(self.W, 1, idx)
            old_w = ww @ v.clone().detach().unsqueeze(1).float()
...