[REDO] Parallel: More precise boundary in ObjectStartArray::object_starts_in_range

The fix in JDK-8279699 uses `align_up` for `end_addr`, which leads to unreserved space if the result, after align_up, is reserve_end.
(The original code on master should suffer from the same problem as well if the passed-in arg `end_addr` is `reserve_end`.)

One way to workaround the problem could be:

```
  if (start_addr == end_addr) {
    // empty
    return false;
  }
  ...
  jbyte* end_block = block_for_addr(end_addr - 1);

  for (jbyte* block = start_block; block <= end_block; block++) {
```