fix: symbols / sections followups

[martin-olivier]

Follow-ups from a review pass over the codebase after the section collection feature landed. One commit per issue to make review easier.

Bug fix

symbols() returned only functions on Linux (bc7d2ec)

The ELF symbol collection filtered on STT_FUNC, while Windows (export directory) and macOS (LC_SYMTAB) both return variables too. As a result get_symbol() could never resolve a namespaced global variable through its demangled name fallback on Linux, even though the README documents exactly that usage.

Variables at global scope were unaffected, because the Itanium ABI leaves them unmangled and dlsym resolved them directly, which is why no test caught it. The example/ program hits this: it calls get_variable("example::magic").

Now collects STT_OBJECT as well, with a namespaced variable added to the test library.

Other changes

Commit	Change
203b6d5	get_symbol() re-demangled every symbol during its fallback, although symbols() already stores a demangled_name. On MSVC that meant four UnDecorateSymbolName calls per symbol instead of two.
2846138	Remove a stray unused <iostream> include sitting mid-file in the tests.
2773130	The version was hardcoded in six places with nothing keeping them in sync. dylib.hpp now exposes DYLIB_VERSION_MAJOR/MINOR/PATCH and CMake parses them from the header. The macros also let users feature-detect an API at compile time. Bumped to 3.1.0, since sections(), collection_error and section_collection_error are new public API.
41b8dab	native_handle() is now const, like every other accessor.
29e68c8	The ELF symbol table size is deduced from the DT_SYMTAB/DT_STRTAB distance, which assumes an ordering the spec does not guarantee. The reverse layout underflowed into a huge unsigned value. Bails out instead, and iterates with a matching unsigned counter.
aa3856a	The POSIX constructor held a file descriptor for the whole library lifetime, though it is only needed by symbols() on macOS and sections() on macOS and Linux. Now opened through an RAII guard only while the file is read, so a process loading many plugins no longer pays a descriptor per library.
4b50716	The example was never built by CI and fetched a released tag rather than the local tree, so nothing kept it working. A new job overrides the fetch with FETCHCONTENT_SOURCE_DIR_DYLIB and builds/runs it against the branch on all three platforms.

Behavior change worth noting

aa3856a means the constructor no longer reports an unreadable library file. That error now surfaces from symbols() or sections(), the operations that actually need to read it.

Validation

Locally on macOS: clean build with DYLIB_WARNING_AS_ERRORS=ON across C++11/14/17/20, all 19 tests pass, clang-format and clang-tidy clean, and the example builds and runs against the local tree. The descriptor change was verified by inspecting /dev/fd before and after symbols()/sections().

The Linux and Windows legs are unverified locally and rely on CI. The new Windows example job in particular is unproven; if it trips, the likely cause is the binary path or the DLL-next-to-exe assumption rather than library code.