[Bugfix] Resolve RDMA NUMA location after memory registration

[jacklin78911-collab]

Description

Untouched anonymous host memory has no NUMA node before registration. Resolving its location at that point records * in the RDMA buffer metadata, allowing NIC selection across unrelated NUMA nodes even after registration pins the pages. This is the registration-order regression reported in #4322.

Resolve the original buffer's first-page location after the first chunk's MRs are registered, then reuse that location for all chunks. Explicit locations, registration parallelism and transfer-path selection are unchanged. A failed NUMA query retains the existing * fallback; it does not trigger registration rollback.

This addresses #4322 without a new placement policy or pre-touch pass. #3657 concerns Store allocation placement; #1974 concerns GPU/PCI topology ranking; neither fixes this registration ordering.

Module

• Transfer Engine (mooncake-transfer-engine)

Type of Change

• Bug fix

How Has This Been Tested?

Linux CPU build from 16b7ba3c, USE_CUDA=OFF, USE_TENT=OFF:

cmake --build build --target rdma_memory_location_test rdma_context_reprobe_test rdma_buffer_range_index_test rdma_transport_submit_task_test -j4
ctest --test-dir build --output-on-failure --no-tests=error \
  -R '^(rdma_memory_location_test|rdma_context_reprobe_test|rdma_buffer_range_index_test|rdma_transport_submit_task_test)$'

• The new test interposes the verbs/NUMA boundary and exercises the real registration, MR-map and metadata code. Untouched single-MR and chunked buffers publish * on the original code and cpu:1 with this fix.
• All four CTest targets pass. The new test also covers explicit location preservation and rollback after a later chunk fails.
• Hardware throughput recovery has not been measured here; the dual-host performance numbers in [Bug]: Host RDMA regression from resolving NUMA before memory registration (0.3.13.post1) #4322 belong to the issue reporter.

Checklist

• I have performed a self-review of my own code (human review pending; draft)
• I have formatted my code using ./scripts/code_format.sh
• I have run pre-commit on the files changed in this PR and all hooks pass
• I have added tests to prove my changes are effective
• Documentation / RFC: not applicable; no API or configuration change.

AI Assistance Disclosure

• AI tools were used

[staryxchen]

LGTM. @stmatengss please double-check, thx~

[staryxchen]

getMemoryLocation() never returns an empty vector. A missing NUMA node and a failed numa_move_pages both come back as one entry with location *, so this entries.empty() branch cannot run. On a real query failure the code publishes * and does not roll back.
Delete the empty check and use entries[0].location directly. LaterRegistrationFailureStillRollsBack covers a later ibv_reg_mr failure, not this branch.