Add PyTorch's core scaled dot-product attention (SDPA) to Optimus. This
automatically uses flash attention 2, or memory efficient attention, if
the hardware supports it. If it doesn't, falls back to manual
implementation.

Training should be much faster with this; memory should also be around
half what it was before.