Multidevice Executor add accessor to fusion executor caches

[cowanmeg]

Add MultiDeviceExecutor::getFusionExecutorCaches which returns a vector of pointers to the multidevice executor's fusion executor caches.

Note, Also renames the field workspace to workspace_ to match naming conventions.

cc @Priya2698 for #2386

[cowanmeg]

!build

[samnordmann]

LGTM! One thing: I'm wondering how the caller of getFusionExecutorCaches can figure out what is the order of the FusionExecutorCache stored in the vector without accessing workspace_.group_run_order. Is my comment relevant?

[cowanmeg]

LGTM! One thing: I'm wondering how the caller of getFusionExecutorCaches can figure out what is the order of the FusionExecutorCache stored in the vector without accessing workspace_.group_run_order. Is my comment relevant?

Yup relevant question. I think the only practical use case is for simple tests like these matmul tests where we have a linear sequence of computation so we know the order. The background for this PR is for matmul with ATen execution we want to check that the executor actually goes through ATen (#2386 (comment)) but to do that we need to access the FusionExecutorCache.

[Priya2698]

Thanks for this PR -- I added these checks to the multidevice matmul tests!