DEVICE/API: Change addresses to offsets and add channel_id

[michal-shalev]

What?

Implement support for passing local and remote addresses to UCX device memory lists, enabling GPU-initiated transfers with pre-configured address pairs and offset-based operations, and add channel ID to the APIs.

Why?

Improve user experience by allowing GPU kernels to work with simple offsets relative to pre-configured base addresses, rather than requiring users to manage and pass absolute addresses for every transfer operation.
Adding channel ID to the device APIs would enable pre-resolved QP/channel selection at runtime, allowing users to register memory once for all experts and then dynamically choose the correct channel during execution, which reduces connection setup time and initialization overhead.

How?

• Extended createGpuXferReq to accept local memory objects, remote rkeys, and remote addresses
• Added getBase() method to nixlUcxMem to access local addresses
• Added getSize() method to nixlUcxMem to access descriptor size
• Modified createGpuXferReq implementation to populate local_addr and remote_addr fields in UCX memory list elements
• Updated UCX backend to extract addresses from descriptors and pass them to the device memory list creation
• Added channel_id to NIXL device APIs

[github-actions]

👋 Hi michal-shalev! Thank you for contributing to ai-dynamo/nixl.

Your PR reviewers will review your contribution then trigger the CI to test your changes.

🚀

[itayalroy]

This PR is missing a change to nixlGpuSignal:

struct nixlGpuSignal {
    uint64_t inc = 0;
    uint64_t remote_addr = 0;
};

should be changed to:

struct nixlGpuSignal {
    uint64_t inc = 0;
    uint64_t offset = 0;
};

[michal-shalev]

/build