fix(io): build under Emscripten without per-consumer workarounds

Two independent problems stopped an Emscripten build of this library, and both forced every
consumer to know something it should not have to.

RandomAccess.cpp: all twelve calls to ThrowNative() sit in a _WIN32 or POSIX branch, and each of
those three-way #if blocks answers Emscripten with PlatformNotSupportedException instead. On that
target the helper genuinely has no caller. GCC does not diagnose that; Clang, which Emscripten uses,
reports "unused function 'ThrowNative' [-Werror,-Wunused-function]" and the build stops. It is now
compiled only where it is used -- no suppression, and the dead code stays out of the wasm binary.

NativeDetail() had to move inside the same guard: its only caller is ThrowNative, so excluding just
ThrowNative moves the identical error one line up.

modules/io-compression: Emscripten ships zlib as a PORT rather than a system library, so
find_package(ZLIB REQUIRED) failed with "missing: ZLIB_LIBRARY ZLIB_INCLUDE_DIR" until the consumer
ran `embuilder build zlib` and passed both paths by hand. The module now requests the port itself
when targeting Emscripten and keeps find_package everywhere else.

That link option is PUBLIC deliberately. This is a STATIC library, so -sUSE_ZLIB=1 has to reach the
final executable's link line too; PRIVATE would configure and compile cleanly and then fail whoever
links the result, which is the worse of the two failures.

Verified with emsdk 6.0.6 against the CNA consumer, with NO workaround flags of any kind:
  * emcmake configure succeeds -- no "Could NOT find ZLIB"
  * sharp_runtime_io compiles (this was the -Werror failure)
  * sharp_runtime_io_compression compiles and links

Native is unaffected -- the guard keeps both helpers, and the CMake else() branch is unchanged. The
CNA HEADLESS suite is identical before and after: 6389 tests, 6172 passed, 217 skipped, 0 failed.

Reported as finding 3 in openeggbert/cna's threeissues.md, which is being updated to match.

Author: robertvokac

modules/io-compression/CMakeLists.txt

@@ -2,8 +2,21 @@
 # Copyright (c) Robert Vokac and contributors
 
 function(_sharp_runtime_setup_io_compression target)
-    find_package(ZLIB REQUIRED)
-    target_link_libraries("${target}" PRIVATE ZLIB::ZLIB)
+    if(EMSCRIPTEN)
+        # Emscripten ships zlib as a PORT, not as a system library, so find_package(ZLIB) fails
+        # with "missing: ZLIB_LIBRARY ZLIB_INCLUDE_DIR" until someone runs `embuilder build zlib`
+        # and passes both paths by hand. Requesting the port here means every consumer no longer
+        # has to know that.
+        #
+        # The link option is PUBLIC on purpose: this is a STATIC library, so the port has to reach
+        # the FINAL executable's link line as well. PRIVATE would configure and build cleanly and
+        # then fail whoever links the result, which is the worse failure of the two.
+        target_compile_options("${target}" PUBLIC "-sUSE_ZLIB=1")
+        target_link_options("${target}" PUBLIC "-sUSE_ZLIB=1")
+    else()
+        find_package(ZLIB REQUIRED)
+        target_link_libraries("${target}" PRIVATE ZLIB::ZLIB)
+    endif()
 endfunction()
 
 sharp_runtime_register_module(

modules/io/src/System/IO/RandomAccess.cpp

@@ -80,6 +80,15 @@ intcs CheckedVectorCount(std::size_t size) {
     return static_cast<intcs>(size);
 }
 
+// Every caller of the two helpers below sits in a _WIN32 or POSIX branch, and each of those
+// three-way #if blocks answers Emscripten with PlatformNotSupportedException instead. So on that
+// target the helpers genuinely have no caller, and Clang -- which Emscripten uses, where GCC does
+// not diagnose this -- reports "unused function 'ThrowNative'" against -Werror and the build stops.
+// Compiling them only where they are used fixes it without a suppression, and keeps the dead code
+// out of the wasm binary. NativeDetail has exactly one caller, ThrowNative, so both must be inside
+// the same guard: excluding only ThrowNative moves the same error one line up.
+#if !defined(__EMSCRIPTEN__)
+
 /// The native reason a syscall failed, appended to the IOException message. Before #2100 every
 /// failure surfaced as a bare "RandomAccess::<member> failed" with the reason discarded, so a
 /// caller could not tell EBADF from EINVAL from ENOSPC.
@@ -94,6 +103,8 @@ std::string NativeDetail(int code) {
     throw IOException(std::string(member) + " failed: " + NativeDetail(code));
 }
 
+#endif // !defined(__EMSCRIPTEN__)
+
 } // namespace
 
 int64_t RandomAccess::GetLength(int fd) {