Adds settings.trace_mode and skip cholesky jitter

[feynmanliang]

We would like to use gpytorch in our NUTS sampler. While the linked code works, it does not utilize JIT compilation (use_nnc) because of control flow.

I understand trace_mode was originally used to disable custom pytorch functions back when JIT did not support compiling through them. This PR (along with cornellius-gp/gpytorch#2163) revives the setting for the AoT JIT compiler (https://pytorch.org/functorch/nightly/aot_autograd.html).

• Moves settings.trace_mode from gpytorch to linear_operators
• Skips conditional control flow when trace_mode is true, currently:
  • PSD check / jitter in linear_operators.utils.cholesky
  • Diagonal x1_eq_x2 check in gpytorch.kernels.kernel

[gpleiss]

This makes sense to me! It seems like we'd want a PR in GPyTorch to ensure that gpytorch.settings.trace_mode also calls linear_operator.settings.trace_mode.

[feynmanliang]

Sounds good, I will send a commit for the unit tests soon.