Individual COMMANDS in define_custom_command are not guaranteed to
execute sequentially, meaning that the inline headers may be broken if
doing a parallel build with make. Each command is extracted into its
own custom_command (some are merged, such as subsequent sed calls) so
that CMake can properly make sense of the dependencies and generate
consistent headers in parallel.

This requires minimum CMake 3.18, which is not particularly
recent. This provides a better cross-platform way of concatenating
files.

Commands in the list of SLEEF_HEADER_COMMANDS were sometimes observed
to overwrite each other when doing a parallel build. Instead of
building a list of commands, build a list of temporary files to depend
on, add a custom_command for each of them, and cat them all together
at the end.

Also added a helper for concatenating files, as this is becoming quite
a common operation.

Refactored out some shared logic - there is likely more to be done
here.

By setting a few optional params the updated header build can be
reused for CUDA - manually verified that the headers are identical.

Individual COMMANDS in define_custom_command are not guaranteed to
execute sequentially, meaning that the inline headers may be broken if
doing a parallel build with make. Each command is extracted into its
own custom_command (some are merged, such as subsequent sed calls) so
that CMake can properly make sense of the dependencies and generate
consistent headers in parallel.

Reusing the commands for the inline headers for the CUDA header, and
use concat_files for disp files.

Now that custom_commands are atomic, parallel builds should be
supported with make.

This was mistakenly removed, now added back.