perf: optimize large EPUB indexing from O(n²) to O(n log n)

Three optimizations for EPUBs with many chapters (e.g. 2768 chapters):

1. OPF idref→href lookup: Build sorted hash index during manifest parsing,
   use binary search during spine resolution. Reduces ~4min to ~30-60s.

2. TOC href→spineIndex lookup: Build sorted hash index in beginTocPass(),
   use binary search in createTocEntry(). Reduces ~4min to ~30-60s.

3. ZIP central-dir cursor: Resume scanning from last position instead of
   restarting from beginning. Reduces ~8min to ~1-3min.

All optimizations only activate for large EPUBs (≥400 spine items).
Small books use unchanged code paths.

Memory impact: ~33KB + ~39KB temporary during indexing, freed after.
Expected total: ~17min → ~3-5min for Shadow Slave (2768 chapters).

Also adds phase timing logs for performance measurement.

lib/ZipFile/ZipFile.h

@@ -25,6 +25,10 @@ class ZipFile {
   ZipDetails zipDetails = {0, 0, false};
   std::unordered_map<std::string, FileStatSlim> fileStatSlimCache;
 
+  // Cursor for sequential central-dir scanning optimization
+  uint32_t lastCentralDirPos = 0;
+  bool lastCentralDirPosValid = false;
+
   bool loadFileStatSlim(const char* filename, FileStatSlim* fileStat);
   long getDataOffset(const FileStatSlim& fileStat);
   bool loadZipDetails();