Use second order AD directly

[rsenne]

Currently to calculate an HVP we use first-order AD on a supplied gradient. Right now two things are preventing the use of second order AD in this package:

1. Batching isn't available in DI
2. I haven't been able to get second order AD on the GPU working with any package (haven't tried a few so this is non-exhaustive).

Issue 2 is obviously a bigger block on this integration so, I will make some MWEs and file them upstream where necessary.

[rsenne]

@wsmoses do you want me to open issues related to this upstream as well? I think its related to the last issue i opened in that I hit a similar error message. So I can also add it there just LMK