Speed up tests with federated learning enabled

[rongou]

Fixes #8334
Fixes #8338

Enabling federated learning slowed down some cpu-only tests, mainly because we increased the number of symbols drastically in libxgboost.so, and the XGBoostAPIGuard throwing an exception and ignoring it is very expensive with symbol lookups. Fixes:

• When federated learning is enabled, pass -Wl,--exclude-libs,ALL to the linker to ignore symbols in .a libraries (see https://sourceware.org/binutils/docs/ld/Options.html).
• Don't throw exception and then ignore it in XGBoostAPIGuard.
• Hide c++ symbols when building gRPC (see https://gcc.gnu.org/wiki/Visibility).

Final result running python/test_updaters.py::TestTreeMethod::test_categorical_ohe:

• Federated learning disabled: 14.33 seconds
• Federated learning enabled: 13.99 seconds

Thanks to @eordentlich for finding the root cause.

@trivialfis @hcho3

[hcho3]

LGTM, assuming the CI passes. Thanks a lot for fixing it so quickly!

[trivialfis]

Thank you for the fix! That is quick! Please help set the enable parameter to True:

xgboost/python-package/xgboost/testing.py

Line 44 in 031d66e

def timeout(sec: int, *args: Any, enable: bool = False, **kwargs: Any) -> Any:

.

[hcho3]

The CPU tests are now very fast:

[hcho3]

@rongou @eordentlich The 1.7 release is unblocked now. Thanks!