[BUG]: relocation R_X86_64_TPOFF32 against hidden symbol `_ZZN8pybind116handle15inc_ref_counterEmE7counter' can not be used when making a shared object

Required prerequisites

• Make sure you've read the documentation. Your issue may be addressed there.
• Search the issue tracker and Discussions to verify that this hasn't already been reported. +1 or comment there if it has.
• Consider asking first in the Gitter chat room or in a Discussion.

What version (or hash if on master) of pybind11 are you using?

2.10.1

Problem description

/usr/bin/ld: build/temp.linux-x86_64-3.10/csrc/transformer/ds_transformer_cuda.o: relocation R_X86_64_TPOFF32 against hidden symbol `_ZZN8pybind116handle15inc_ref_counterEmE7counter' can not be used when making a shared object

Reproducible example code

I was following https://www.deepspeed.ai/tutorials/advanced-install/.

Specifically, it can be reproduced with

`DS_BUILD_TRANSFORMER_INFERENCE=1 pip install .`

The command that failed is 


root@4212f9fa521e:/app# x86_64-linux-gnu-g++ -shared -Wl,-O1 -Wl,-Bsymbolic-functions -Wl,-Bsymbolic-functions -g -fwrapv -O2 -Wl,-Bsymbolic-functions -g -fwrapv -O2 -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 build/temp.linux-x86_64-3.10/csrc/transformer/cublas_wrappers.o build/temp.linux-x86_64-3.10/csrc/transformer/dropout_kernels.o build/temp.linux-x86_64-3.10/csrc/transformer/ds_transformer_cuda.o build/temp.linux-x86_64-3.10/csrc/transformer/gelu_kernels.o build/temp.linux-x86_64-3.10/csrc/transformer/general_kernels.o build/temp.linux-x86_64-3.10/csrc/transformer/normalize_kernels.o build/temp.linux-x86_64-3.10/csrc/transformer/softmax_kernels.o build/temp.linux-x86_64-3.10/csrc/transformer/transform_kernels.o -L/usr/local/lib/python3.10/dist-packages/torch/lib -L/usr/local/cuda/lib64 -lc10 -ltorch -ltorch_cpu -ltorch_python -lcudart -lc10_cuda -ltorch_cuda -o build/lib.linux-x86_64-3.10/deepspeed/ops/transformer/stochastic_transformer_op.cpython-310-x86_64-linux-gnu.so

### Is this a regression? Put the last known working version here if it is.

Not a regression

Start with the reported DS_BUILD_TRANSFORMER_INFERENCE=1 pip install . reproduction and inspect ds_transformer_cuda.o together with the shown shared-object link command. Trace the _ZZN8pybind116handle15inc_ref_counterEmE7counter relocation and compare the build behavior across relevant configurations; done means the shared object links without the reported error.