[Jiayi-Wang-db]

Rationale for this change

Support flexible data layouts and statistics over column-chunk, page, or arbitrary row ranges, while allowing readers to fetch placement and statistics independently.

What changes are included in this PR?

ModularFooter.thrift defines the footer directory and its metadata streams:

• Typed Parquet page streams located by offset, length, and codec.
• Shared column-to-chunk and chunk-to-page indexes, allowing each column to have a different chunk count.
• Direct chunk offsets and sizes, a dictionary table with per-chunk references, and optional row-group and page metadata.
• Tagged row-range statistics with exact null, NaN, and value counts and typed min/max bounds. Chunk and page statistics reuse shared row positions.
• PME structures for directory authentication, shared-stream protection, and separately encrypted column statistics.
• A 4-byte little-endian directory length followed by the 4-byte magic; the length includes any directory signature.

The definition uses the ModularFooter root and the parquet.modular / org.apache.parquet.format.modular namespaces. Format-version and PME module-type allocations remain open.

Do these changes have PoC implementations?

The design and measurements describe the prototype readers and writers. Measurements cover plaintext metadata streams and pruning predicates; exact-count statistics, this directory serialization, encryption, and column sets are outside the measured scope. The linked design describes column sets as an optional extension for very wide files.

Validation: the definition parses, and Apache Thrift 0.24.0 generates Java and C++ bindings; the generated C++ passes a syntax check, and all 83 generated Java classes compile with the cached Thrift 0.23.0 runtime. Full Maven verification could not complete locally because dependency repositories were unavailable.

Related to #622.