New allreduce algo for small message size

[Binyang2014]

New algo for message size < 32KB, command: mpirun --allow-run-as-root -tag-output -np 8 -x LD_PRELOAD=/root/mscclpp/build/apps/nccl/libmscclpp_nccl.so -x MSCCLPP_DISABLE_CHANNEL_CACHE=1 ./build/all_reduce_perf -b 1K -e 32K -f 2 -c 1 -G 1 -n 100 -d half
Tested on H100
Perf:

[1,0]<stdout>:#                                                              out-of-place                       in-place          
[1,0]<stdout>:#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
[1,0]<stdout>:#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)       
[1,0]<stdout>:        1024           512      half     sum      -1     4.65    0.22    0.39      0     4.60    0.22    0.39      0
[1,0]<stdout>:        2048          1024      half     sum      -1     4.96    0.41    0.72      0     4.93    0.42    0.73      0
[1,0]<stdout>:        4096          2048      half     sum      -1     5.12    0.80    1.40      0     5.12    0.80    1.40      0
[1,0]<stdout>:        8192          4096      half     sum      -1     5.11    1.60    2.81      0     5.08    1.61    2.82      0
[1,0]<stdout>:       16384          8192      half     sum      -1     5.47    3.00    5.24      0     5.44    3.01    5.27      0
[1,0]<stdout>:       32768         16384      half     sum      -1     6.24    5.25    9.19      0     6.28    5.22    9.14      0
[1,0]<stdout>:# Out of bounds values : 0 OK
[1,0]<stdout>:# Avg bus bandwidth    : 3.29145

Old:

[1,0]<stdout>:#                                                              out-of-place                       in-place          
[1,0]<stdout>:#       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
[1,0]<stdout>:#        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)       
[1,0]<stdout>:        1024           512      half     sum      -1     5.02    0.20    0.36      0     5.12    0.20    0.35      0
[1,0]<stdout>:        2048          1024      half     sum      -1     5.28    0.39    0.68      0     5.29    0.39    0.68      0
[1,0]<stdout>:        4096          2048      half     sum      -1     5.45    0.75    1.32      0     5.46    0.75    1.31      0
[1,0]<stdout>:        8192          4096      half     sum      -1     5.50    1.49    2.61      0     5.51    1.49    2.60      0
[1,0]<stdout>:       16384          8192      half     sum      -1     5.79    2.83    4.95      0     5.80    2.82    4.94      0
[1,0]<stdout>:       32768         16384      half     sum      -1     7.36    4.45    7.79      0     7.36    4.46    7.80      0
[1,0]<stdout>:# Out of bounds values : 0 OK
[1,0]<stdout>:# Avg bus bandwidth    : 2.94887 

[Copilot]

Pull Request Overview

This PR introduces a new allreduce algorithm optimized for small message sizes (< 32KB) using NVLS packet-based communication, improving bandwidth performance by approximately 9% compared to the previous implementation.

Key changes:

• New AllreduceNvlsPacket algorithm specifically for messages ≤ 32KB
• Enhanced SwitchChannelDeviceHandle with packet-based broadcast functionality
• Updated algorithm selection logic to prioritize the new packet algorithm for small messages

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 6 comments.

File	Description
include/mscclpp/switch_channel_device.hpp	Added packet broadcast specialization and multimem store for LLPacket
apps/nccl/src/nccl.cu	Integrated new algorithm, optimized selection logic with static variables, fixed control flow
apps/nccl/src/allreduce.hpp	Declared AllreduceNvlsPacket class and allreduceNvlsPacket kernel
apps/nccl/src/allreduce.cu	Implemented AllreduceNvlsPacket algorithm with adapter and kernel launch logic

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[caiomcbr]

LGTM