cmake: make the installed package consumable

[petlenz]

Fixes #18 — and a second defect found while verifying the first.

Both were invisible to every in-tree build, because those use BUILD_INTERFACE, which was always correct.

1. The exported target had no include directory

include(GNUInstallDirs) sat 18 lines after the $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}> that depends on it. ${} expands when the line is read, so the variable was empty and the generator expression collapsed to $<INSTALL_INTERFACE:>:

set_target_properties(numsim-core::numsim-core PROPERTIES
  INTERFACE_COMPILE_FEATURES "cxx_std_23"
)

Consumers could find the package and not compile against it. An empty INSTALL_INTERFACE is legal, so nothing warned. The install(FILES ... DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/...) call is after the include, which is why the headers landed correctly and only the export was wrong.

2. Headers installed flattened, and most were not installed at all

install(FILES) puts everything in one directory, so include/numsim-core/property_graph/*.h landed beside the top-level headers and #include <numsim-core/property_graph/property_traits.h> could not resolve. It also installed only the files listed in ${PROJECT_NAME}_HEADER — 13 of the 23 headers in the tree.

before:  1 directory,  13 headers
after:   2 directories, 23 headers

install(DIRECTORY ... FILES_MATCHING PATTERN "*.h") installs all of them and preserves the structure. It also stops the list going stale: a new header no longer has to be added in two places.

Verified end to end

A three-line external project that calls find_package(numsim-core) and includes a property_graph header now configures, compiles and runs. Before this it failed at configure on the missing include directory, and past that on the missing subdirectory.

Checked against the downstream case too: numsim-materials had its own separate packaging defect (its generated Config re-found none of its dependencies). With that fixed and these two, a consumer builds against an installed numsim-materials for the first time.

Worth considering separately

Neither project's CI runs install or builds a consumer against the result, which is why all of this sat behind permanently green builds. A CI step that installs and compiles a three-line consumer would have caught every one of these.

[petlenz]

Critical review. One finding, fixed in the branch.

${PROJECT_NAME}_HEADER was left set at line 30 and referenced nowhere after install(DIRECTORY) replaced its only use — dead code, and specifically the dead code whose staleness caused half this bug. A hand-maintained header list that must be edited whenever a file is added is exactly what installed 13 of 23. Removed.

What I checked and could not break:

• install(DIRECTORY ... FILES_MATCHING PATTERN "*.h") installs 23 headers across 2 directories and pulls in nothing unintended — the tree contains only .h files.
• The exported target carries INTERFACE_INCLUDE_DIRECTORIES "\${_IMPORT_PREFIX}/include" after the GNUInstallDirs hoist.
• End to end: an external project calling find_package(numsim-core) and including a property_graph header configures, compiles and runs.