Add python wrapper for system memory resource

[rongou]

Description

Follow up on #1581 to add access to the system memory resource in python.

Fixes #1622

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[rongou]

@wence- @vyasr can you take a look? Thanks!

[rongou]

ping @harrism @wence- @vyasr

[vyasr]

Apologies, Lawrence and are swamped with multiple large P0 efforts going into this release. I'll try to get to this today.

[harrism]

Looks fine, but I question whether we want to add more options to initialize.

[harrism]

Thanks for compromising, @rongou ! As part of our upcoming Python refactoring, we should consider how to make it easier to handle setting MRs more generically.

[bdice]

An adaptor accepts another resource as an argument (e.g. a logging adaptor adds logging, but passes through to the underlying MR). This is not an adaptor, it is implemented like an entirely new resource.

Adaptors must be composable. It cannot be specific to the System MR and must accept other MRs if it is to be an adaptor.

[bdice]

Should we redesign this so that the SystemMemoryResource has an argument headroom and we don't need the "adaptor" at all? Or rename this so it's not an adaptor?

[rongou]

The C++ side has both the system mr and the adaptor, but I guess we don't have to strictly mirror it in python. I'd be happy to have a single SystemMemoryResource with an argument.

[rongou]

Done.

[vyasr]

I think there's room for future improvement here, but we definitely can't do that in this PR so I don't want to hold things up for this release.

[vyasr]

Approving again post the final split of classes on the Python side.

[harrism]

/merge