Improve BPCells compatibility (JackStraw + RunPCA)

[bnprks]

These are some changes brought about from this BPCells issue, then referenced in Seurat issue #8267

1. Make the JackStraw function compatible with BPCells objects. This approach shuffles the matrix subset as a dense matrix, but leaves the remainder of the data on-disk. For best performance the BPCells object should be stored in row-major order so that the matrix subsets can be performed efficiently, but it will work either way.
2. Improve performance of RunPCA by using the BPCells svds function. This takes advantage of the same solver RSpectra uses, but at the C++ level to avoid overheads of repeatedly creating BPCells C++ objects on each solver iteration. It also avoids the pitfall that irlba can only use its fast C version with in-memory matrices and otherwise falls back to a slower pure-R implemenation.