Background

The NetworKit binary graph format (.nkb) is positional and node-ID based:
the reader decodes one entry per node ID u ∈ [0, header.nodes).
However, the writer previously serialized data only for existing nodes, which caused stream misalignment when graphs contained deleted nodes (non-continuous IDs). This could result in assertions and crashes when reading such files (see #1278).

Summary of changes

1. Use upperNodeIdBound() instead of numberOfNodes()
numberOfNodes() counts only existing nodes, but the binary format must cover the entire node-ID space, including holes. Using upperNodeIdBound() ensures writer and reader agree on the node range.

2. Iterate over node IDs instead of G.forNodes(...)
G.forNodes(...) skips deleted node IDs.
   All writer loops that emit per-node positional data must therefore iterate over node IDs and explicitly encode deleted nodes as zero-degree entries. This applies to:

   • base node flags
   • adjacency lists
   • transpose lists
   • weights
   • edge IDs

3. Explicitly handle deleted nodes during chunk size computation
   Even deleted nodes must contribute a degree entry (0) so that offset tables match the actual byte layout.
   Without this, adjacency data is decoded at incorrect positions.

4. Fix inverted DELETED_BIT semantics
   The previous implementation had inverted semantics for DELETED_BIT in both writer and reader:

   • the writer set DELETED_BIT for existing nodes
   • the reader removed nodes when DELETED_BIT was not set
     These two inversions accidentally canceled out for dense graphs, masking the bug. This PR fixes the semantics to match the
     name and intended meaning:
   • the writer sets DELETED_BIT only for deleted nodes
   • the reader removes nodes when DELETED_BIT is set

Result

• Existing tests pass, including testWriteReadNonContinuous and testWriteReadNonContinuousDirected mentioned in NetworkitBinaryReader/Writer + deleted nodes + deleted edges #1278.

Fixes #1278