[codex] Add source materialization cache and nanoarrow backend

[ashione]

What changed

• add a generic source materialization cache layer keyed by source format, options, absolute path, file size, and mtime
• teach DataflowSession::read_csv(...) to reuse persisted materialized data instead of reparsing unchanged sources
• add a nanoarrow_ipc materialization backend alongside the internal binary row batch backend
• add native C++ regression coverage for source materialization roundtrip and cache reuse
• add a Python regression test that executes SQL after reading a 20k-row CSV, then verifies a second session reuses the existing materialized source file

Why

Repeated reads of the same formatted source were always going back to the original file. This branch adds a reusable, cross-session materialization path so follow-up queries can skip the original source parse when the source fingerprint is unchanged.

Impact

• batch read_csv(...) can now reuse persisted source materialization across sessions
• the materialization layer is no longer CSV-specific, which leaves room for other formatted sources to use the same mechanism
• nanoarrow_ipc is available as a persisted data format without requiring the full Arrow C++ dependency tree

Validation

• bazel test //:source_materialization_test
• bazel build //:velaria_pyext
• bazel test //python_api:source_materialization_test