Make CUDA lineinfo optional.

[trivialfis]

Ref #11604 .

Before:

$ strip libxgboost.so 
$ du -h libxgboost.so 
445M    libxgboost.so

After:

$ strip libxgboost.so 
$ du -h libxgboost.so 
148M    libxgboost.so

[trivialfis]

cc @hcho3

[hcho3]

Can we create a new CMake option to enable the lineinfo? Or is it redundant given the USE_DEVICE_DEBUG option?

[trivialfis]

It's redundant when debug mode is enabled.

After this PR , it will be enabled when nvtx is enabled since the lineinfo is most relevant for profiling.

Please note that, it's not necessary to enable nvtx for profiling. I'm just grouping them under one flag.

We can remove this nvcc flag entirely and rely on the user to supply a cmake compiler flag. I don't have a strong opinion on this choice, remove the flag, split up the flag, or group the flag under nvtx.

[hcho3]

I ran some experiments on my end.

CMake options: -DUSE_DLOPEN_NCCL=ON -DPLUGIN_FEDERATED=OFF (all SMs enabled)

Before (e4406da):

• Uncompressed: 832 MB
• Compressed: 572 MB

After (dcf8afe):

• Uncompressed: 257 MB (70% savings)
• Compressed: 167 MB (70% savings)