Fix bugs and improve performance

[sldyns]

1. Resolved the issue of continuous GPU memory growth by altering weights = attn to weights = attn.detach().
2. Enhanced execution speed, decreasing memory consumption, by substituting the masked Multi-Layer Perceptron (MLP) operation within FeatureEmbed with sparse matrix multiplication via the creation of a new SparseLinear class.
3. No need for torch==1.7.1, it works perfectly with torch==2.1.0.