fix(parquet): cache deferred page index loads - #24004

Which issue does this PR close?

• Closes Deferred page-index load in the Parquet opener bypasses FileMetadataCache, causing repeated uncached I/O when the skip heuristic doesn't fire #23978.

Rationale for this change

The Parquet opener intentionally loads footer-only metadata first so row-group
statistics can decide whether page-index I/O is necessary. When the page index
is needed later, however, the deferred byte-range load bypasses
DFParquetMetadata::fetch_metadata, so the augmented metadata was not written
back to FileMetadataCache. Reopening the same file therefore repeated the
page-index request.

What changes are included in this PR?

• Add a default no-op ParquetFileReaderFactory::cache_metadata hook so
  factories with a metadata cache can retain metadata augmented after the
  initial request without requiring changes from existing implementations.
• Have CachedParquetFileReaderFactory replace the existing footer-only cache
  entry after a deferred page-index load. As with the existing metadata path,
  decrypted metadata is not cached.
• Call the hook after the opener finishes loading the page index.
• Add a regression test that verifies the cache entry gains the page index and
  that a second open performs zero reader byte-range I/O.

Are these changes tested?

Yes.

• cargo test -p datafusion-datasource-parquet (159 passed)
• cargo test -p datafusion-datasource-parquet page_index -- --nocapture
• cargo test -p datafusion-datasource-parquet --all-features test_deferred_page_index_is_cached -- --nocapture
• cargo clippy --all-targets --workspace --features avro,integration-tests,extended_tests -- -D warnings
• cargo clippy -p datafusion-datasource-parquet --all-targets --all-features -- -D warnings
• Full repository formatting, TOML formatting, license, spelling, Prettier,
  workflow-install, and rustdoc checks
• cargo-semver-checks against upstream/main (223 checks passed)

The regression test measures the I/O invariant directly rather than relying on
wall-clock timing, so no separate benchmark was added.

Are there any user-facing changes?

ParquetFileReaderFactory gains a default no-op cache-update hook. Existing
implementations remain source-compatible, and the SemVer check reports that no
version update is required. The default cached reader now avoids repeated
page-index requests when opening the same file.

AI assistance

I used OpenAI Codex while implementing this change. I reviewed the code and
tests and understand the metadata-loading and cache-update path end to end.