Faster dataloader

[dfalbel]

Improve dataloaders performance.

• No longer use coro::generator() in critical parts as it adds a lot of overhead.
• Don't use coro::as_iterator() as this will use <<- from coro, making it slower.
• nit: Don't recreate sliced tensor using Tensor$new() as it adds small overhead and is no longer necessary.
• Vectorized batch samplers allow for faster loading. Specially if using .getbatch instead of .getitem().