Linking is really slow with Ninja CMake generator, but not Unix Makefiles

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?

2.13.1

Problem description

When building a large pybind11 module with GCC+CMake+Ninja, I find that the linking step is serial and takes forever. When the Makefiles generator is specified, linking (LTO) uses all available cpus, and linking is much faster.

environment: Conda-forge, linux x64

GCC 12.3.0
CMake 3.30.0
GNU Make 4.4.1
Ninja 1.12.1

The problem seems related to the fact that Ninja does not support the GNU Make jobserver protocol. Therefore, the -flto flag does serial LTO with Ninja and parallel LTO with Make. You need to explicitly set -flto=auto or -flto=n when you use Ninja for reasonable link times.

It might be a good idea to change set(thin "") to set(thin "=auto") here:
https://github.com/pybind/pybind11/blob/bb05e0810b87e74709d9f4c4545f1f57a1b386f5/tools/pybind11Common.cmake#L327
Clang supports -flto=auto too (with Clang, it means the same as -flto).

Reproducible example code

No response

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

Not a regression

Start at tools/pybind11Common.cmake around line 327 and inspect how the LTO flag is selected for CMake builds. Compare large-module link behavior with the Ninja and Unix Makefiles generators, and consider the reported GCC and Clang behavior. Done means Ninja builds no longer perform serial LTO while Makefiles retain their existing behavior.