`fix`: improve the GPU utilization with better `Tensor` and `Var` handling

[FiberedSkies]

Currently there are a few issues with the way the Matrix, VarMatrix, and Vector wrappers of candle_core::Tensor are being treated, that leads to bad GPU utilization and low cache locality. These include:

1. Frequent Device Transfers: The code creates many small tensors individually, which can cause inefficient GPU memory allocation patterns.
2. Redundant Device/DType Storage: Each Matrix, VarMatrix, and Vector stores its own Device and DType, which is redundant since the underlying Tensor already has this information.
3. Inefficient Small Operations: Operations like creating identity matrices element-by-element are not GPU-optimized.
4. Sequential Processing: The sheaf operations process cells one at a time rather than in batches.