Add get_nvlink_address() to device window API and Triton bindings

[siyengar]

Summary:
Add get_nvlink_address() method to TorchCommWindowNCCLX that returns the
NVLink-mapped device pointer for a peer's window memory. This calls
ncclGetPeerDevicePointer (host-side, NCCLX 2.29+ when available) to resolve
the LSA flat address for NVLink-accessible peers, returning nullptr for peers
not reachable via NVLink. The API guard is set at NCCLX 2.28+ with a nested
2.29+ check for the actual host API call, falling back to returning nullptr
on 2.28.x.

Also add torchcomms_get_nvlink_address() to the Triton device bindings,
which calls ncclGetPeerPointer() (device-side, available in 2.28+) to
compute the NVLink address directly on the GPU.

Changes:

• NcclxApi: Add winGetPeerDevicePointer() virtual method (v2.28+ guard, v2.29+ actual call)
• TorchCommWindowNCCLX: Add get_nvlink_address(peer, offset) method
• torchcomms_device.h/.cu: Add torchcomms_get_nvlink_address extern
• Triton Python wrapper: Add get_nvlink_address() function
• Pybind: Register get_nvlink_address (v2.28+ only)
• Type stubs: Add get_nvlink_address signature

Differential Revision: D95908486

[meta-codesync]

@siyengar has exported this pull request. If you are a Meta employee, you can view the originating Diff in D95908486.