Add tests for partial cudagraph memory and numerics regressions

**Is your feature request related to a problem? Please describe.**
Create tests that can catch regressions for partial cudagraph memory and numerical changes.

- Catching memory regressions is not straightforward since pytorch doesn't provide tooling for tracking the memory usage of graph pools. The current way is to manually comb through the torch memory profile.

- For numerical changes, this is also difficult as pytorch/libraries provide no guarantee cudagraphs call the same kernels as eager mode, meaning without some intervention, graphing over a module will cause numerical changes.

No files or existing tests are named. Start by locating current partial cudagraph and graph-pool coverage, then identify how numerical comparisons are performed between cudagraphs and eager mode. Done means regression tests cover both graph-pool memory behavior and numerical changes.

Written by the indexing model from the issue text.