Several small improvements to the Win32 Cursor implementation:

- Replace double map-lookup and setHandleForZoomLevel() helper with a
  single HashMap.computeIfAbsent() call in win32_getHandle(). The
  helper method contained redundant null/containsKey guards that were
  already enforced by the call-site; both are now unnecessary.

- Change win32_getHandle() return type from boxed Long to primitive
  long, consistent with Font.win32_getHandle(), Image.win32_getHandle()
  and Region.win32_getHandle(). All existing call sites already
  consumed the result as a long, so no callers required changes.

- Fix latent equality bug in equals(): the previous code compared two
  Long wrapper objects with ==, which only gives correct results for
  cached values in [-128, 127]. Cursor handles are OS memory addresses
  and fall well outside that range. With the primitive return type the
  comparison is now a straightforward == on long values, which is both
  correct and avoids boxing.

- Use Java 16+ instanceof pattern matching in equals() to eliminate
  the explicit cast after the type check.

- Remove unnecessary clone() in ImageDataWithMaskCursorHandleProvider
  .validateMask(): the cloned ImageData was only ever read for its
  width/height fields, so the defensive copy was pure overhead.

- Remove redundant 'static' modifier on the nested CursorHandleProvider
  interface declaration; nested interfaces are implicitly static.

- Remove redundant 'final' modifier on the private static method
  getOSCursorIdFromStyle(); private static methods cannot be overridden.

- Remove redundant 'final' modifier on the lambda-captured local
  variable in destroyHandlesExcept(); the variable is effectively final
  without the explicit keyword.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>