Ub transport store dev

[zchuango]

Description

This PR enables the UB (Ultra Band) transport for Mooncake Store, along with several bug fixes and improvements for the UB transport layer in the transfer engine.

Key changes:

1. UB Transport support in Mooncake Store: Added "ub" protocol handling in the store's client service, real client, and utility functions. This includes:

   • UB protocol initialization in Client::InitTransferEngine with device name discovery
   • UbSegmentDeleter for proper cleanup of UB-allocated segments
   • Integration with ub_allocate_memory / ub_free_memory in the store's memory allocation path

2. UB Allocator: Introduced a new ub_allocator module (ub_allocator.h / ub_allocator.cpp) that provides:

   • ub_allocate_memory() — allocates NUMA-local memory via numa_alloc_local and tracks the allocation range
   • ub_free_memory() — frees NUMA-allocated memory and removes the tracked range
   • ub_is_store_memory() — checks if a given address range overlaps with any tracked UB store memory region

3. UrmaEndpoint bug fixes and improvements:

   • Fixed polling logic: moved slices[i] = slice assignment before the success check so that failed slices are also properly returned to the caller
   • Added support for URMA_PORT_ACTIVE_DEFER port state in device open
   • Added error logging when retrieveRemoteSeg fails and when a jetty is not imported
   • Fixed endpoint deletion order: only delete the endpoint after retry count exceeds the maximum (previously it was deleted prematurely on first failure)
   • Set rjetty.tp_type and rjetty.flag fields during connection setup

4. Build & packaging updates:

   • Updated urma dependency to v25.12.0.B081
   • Added ub_allocator.cpp to the UB sources in CMake
   • Added ub_allocator.h to the install headers
   • Added -lurma linkage in p2p-store build script
   • Excluded liburma.so* from auditwheel repair in the wheel build script

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• P2P Store (mooncake-p2p-store)
• Mooncake EP (mooncake-ep)
• Integration (mooncake-integration)
• Python Wheel (mooncake-wheel)
• PyTorch Backend (mooncake-pg)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• New feature
• Bug fix
• Refactor
• Breaking change
• Documentation update
• Other

How Has This Been Tested?

• Verified UB protocol initialization and segment registration in Mooncake Store with UB transport
• Tested ub_allocate_memory / ub_free_memory allocation and deallocation paths
• Validated ub_is_store_memory range overlap detection logic
• Confirmed UrmaEndpoint polling correctly returns both successful and failed slices
• Verified build succeeds with UB transport enabled (including CMake and wheel packaging)

Checklist

• I have performed a self-review of my own code.
• I have formatted my own code using ./scripts/code_format.sh before submitting.
• I have updated the documentation.
• I have added tests to prove my changes are effective.