Think about unitwise fisher, at least for small-ish models

[mmatena]

Why

See if better approximations to the Fisher lead to better merges and larger boosts.

It's possible that this turns out to be more expository than practical, which is fine.

Steps

• Compute number of parameters needed to for unit-wise approximations to see if it is feasible.
• Write the fisher computer.
  • Have the unit-wise approximation stay on the CPU.
  • Perhaps do it in chunks of variables where we save them to disk after computing them.
  • Write the merger.
  • Again, be deliberate in how we load stuff, what resides on the GPU vs CPU, and how we save stuff.