Add optional torch usage for big matrix computations

[bschneidr]

The torch package includes some fast methods for matrix decompositions that can be used as drop-in replacements for the default BLAS/LAPACK matrix computations in the package, with easy-to-use options for GPU computing. This would lead to noticeably improved runtimes for some functions like make_gen_boot_factors() or make_fays_gen_rep_factors(). Let's do the following:

• 1. Add torch to the package suggests
• 2. Add a package-wide option like options(svrep.torch_device = c(NULL, 'cpu', 'cuda'))
• 3. In make_fays_gen_rep_factors(), allow the eigendecomposition to use Torch's eigendecomposition depending on getOption("svrep.torch_device").
• 4. In make_gen_boot_factors(), see if we can generate the multivariate normal draws using torch depending on getOption("svrep.torch_device").