feat(io): fall back RDMA backend to XGMI when no RDMA device is present

[maning00]

Summary

Make IOEngine::CreateBackend(BackendType::RDMA, ...) behave gracefully on hosts without an active RDMA NIC. Previously the call would hit assert(availDevices.size() > 0) deep inside RdmaManager and SIGABRT —
which prevents SGLang's mori PD-disaggregation path (sgl-project/sglang#25094) from running on single-node setups even when XGMI/Infinity Fabric is available between local GPUs.

This PR is the mori-side counterpart: when no RDMA device exists and the operator opts in via MORI_DISABLE_AUTO_XGMI=0, CreateBackend(RDMA) is transparently rerouted to create an XgmiBackend instead. SGLang's existing engine.create_backend(BackendType.RDMA, rdma_cfg) keeps working unchanged on no-RDMA hosts.

Behavior

Scenario	New behavior
Has RDMA device	Unchanged (creates RdmaBackend; EnsureXgmiBackendCreatedIfSupported() still runs)
No RDMA device, default env (MORI_DISABLE_AUTO_XGMI unset / not 0)	throw std::runtime_error("...; set MORI_DISABLE_AUTO_XGMI=0 to enable XGMI-only fallback") (no SIGABRT)
No RDMA device, opt-in (MORI_DISABLE_AUTO_XGMI=0), GPU P2P available	Reroutes to XgmiBackend, engine_desc.port = 1 (sentinel); one-line WARN log
No RDMA device, opt-in, no GPU P2P	throw std::runtime_error("...no usable GPU P2P; cannot create any backend") — startup fail-fast

Implementation notes

• RdmaManager ctor: assert → throw std::runtime_error (assert is compiled out in Release, masking failures).
• RdmaBackend ctor: wrap RdmaContext* in unique_ptr to avoid leak on the new throw path.
• RdmaBackend::CanHandle(): rejects sentinel-port remotes — otherwise a peer's BuildRdmaConn → Connect hits SYSCALL_RETURN_ZERO and exit(-1)s the whole peer process.
• IOEngine::SelectBackend(): every routing path now checks CanHandle(); the previous "first backend wins" tail-fallback is replaced with return nullptr, so unhandled transfers surface ERR_BAD_STATE via the existing SELECT_BACKEND_AND_RETURN_IF_NONE macro.
• Sentinel port 1: must be > 0 (SGLang asserts port > 0); unbindable by non-root, easy to spot in netstat. Defensive checks reject config.port == 1 and any RDMA ephemeral bind that lands on 1.

Test plan

New Case* tests in tests/cpp/io/test_engine.cpp, registered in the cases list. Use MORI_RDMA_DEVICES="__mori_no_such_device_for_test__" to deterministically simulate "no RDMA":

• rdma_backend_has_active_devices_returns_false_when_no_device
• rdma_manager_throws_when_no_active_devices
• create_backend_rdma_throws_by_default_when_no_rdma_device
• create_backend_rdma_falls_back_to_xgmi_when_opted_in
• create_backend_rdma_throws_when_opted_in_but_no_xgmi (no-GPU only)
• explicit_xgmi_then_rdma_without_opt_in_still_throws
• explicit_xgmi_then_rdma_with_opt_in_refreshes_port
• rdma_backend_refuses_sentinel_port_config
• select_backend_returns_null_for_cross_node_under_xgmi_only
• rdma_backend_can_handle_rejects_sentinel_port_remote

Existing RDMA tests guarded by if (!RdmaBackend::HasActiveDevices()) throw TestSkip(...) so the assert → throw change does not break test_engine on no-RDMA CI.

[clintg6]

Finished reviewing and testing the PR. I can confirm this enables the XGMI-only fallback path on nodes with no active RDMA devices when MORI_DISABLE_AUTO_XGMI=0 is set. In SGLang, we can gate that by setting MORI_DISABLE_AUTO_XGMI=0 only when SGLANG_MORI_USE_XGMI=1. I have updated the SGLang PR to let MORI handle XGMI selection. @maning00