Feature: support symmetric heap and unified memory space

[jhchouuu]

Motivation

support symmetric heap and unified memory space

Technical Details

1. support static symmetric heap and unified memory space

   shmem support two style api:
   SymmMemObjPtr + offset:

   ShmemPutMemNbiThread(
       const application::SymmMemObjPtr dest, size_t destOffset,
       const application::RdmaMemoryRegion& source, size_t sourceOffset, size_t bytes, int pe,
       int qpId = 0) 
   

   unified memory :

   ShmemPutMemNbiThread(void* dest, const void* source, size_t bytes,
                                               int pe, int qpId = 0)
   

2. The symmetric heap currently only supports static heaps.
   use env variable MORI_SHMEM_HEAP_SIZE to set default heap memory size

Test Result

test pass:

• mpirun --allow-run-as-root -np 2 ./examples/concurrent_put_thread
• mpirun --allow-run-as-root -np 2 ./examples/concurrent_put_imm_thread
• mpirun --allow-run-as-root -np 2 ./examples/atomic_fetch_thread
• mpirun --allow-run-as-root -np 2 ./examples/atomic_nonfetch_thread
• mpirun --allow-run-as-root -np 2 ./examples/concurrent_put_signal_thread
• torchrun --nnodes=2 --node_rank=0 --nproc_per_node=1 --master_addr=xxx --master_port=1236 examples/ops/dispatch_combine/test_dispatch_combine_internode.py