feat: support gpt-oss in sglang-0.5.9 with clean allocator

[ztang2370]

Summary

Compared to the previous KVGroup approach proposed in #263, this approach replaces the internal KVGroup struct with a multiton pattern, using one FTensorAllocator instance per group_id, lazily created via global_allocator(group_id).

What changed

allocator.hpp: Replaced single g_allocator_ with g_allocators_ map.
allocator.cpp: global_allocator(group_id) lazily creates per-group allocators.
torch_bindings.cpp: Routes group_id to the correct allocator at the binding layer.

Tested gpt-oss-20b on sglang-0.5.9.

[ivanium]

Thanks for the PR! This one does look simpler. QQ: how do we set the kv cache config, such as tensor size, number of layers, etc., into the cpp extension?

[ztang2370]

Thanks for the PR! This one does look simpler. QQ: how do we set the kv cache config, such as tensor size, number of layers, etc., into the cpp extension?

The kv cache config is set when the Python side calls create_kv_tensors(size, dtype_size, dev_str, num_layers, num_kv_buffers, group_id), which is the same entry point as before.
torch_bindings.cpp receives all config + group_id, and calls global_allocator(group_id) to get or lazily create the right allocator instance.
Then in allocator.cpp, create_kv_tensors() stores the config into its own members, creates the zero page, and builds the FTensors.

So each allocator instance gets configured the first time create_kv_tensors is called on it with group_id.

[ivanium]

Oh right I missed that part before. But yeah that makes sense. I am okay with getting the PR in. Thanks for the work!