refactor(parquet-datasource): mechanical cleanup of opener / file_format / row_group_filter

[adriangb]

Which issue does this PR close?

Relates to the discussion in #22024 about the Parquet datasource crate becoming hard to navigate as more features land. Does not close that issue — this is a small, mechanical step that makes future hook/extension work easier.

Rationale for this change

Several files in the parquet datasource crate have grown to bundle multiple responsibilities into one place — making it hard to read and review changes in isolation. This PR is pure code motion: no public API changes (existing import paths preserved via re-exports), no behavior changes.

The hook/extension trait that addresses #22024's broader concern ("the opener is becoming a mini planner") is being prepared as a separate PR so reviewers can evaluate each change on its own merits.

What changes are included in this PR?

Five separate moves grouped into one PR because they're all the same flavor of work:

1. opener.rs (~2,700 LOC) → opener/ module

• New opener/early_stop.rs — EarlyStoppingStream (~107 LOC), the dynamic-filter early-termination wrapper used at the end of build_stream.
• New opener/encryption.rs — EncryptionContext + the ParquetMorselizer::get_encryption_context helpers (~104 LOC). Isolates the #[cfg(feature = "parquet_encryption")] gating that previously bled through the main file.
• New opener/push_decoder_stream.rs — PushDecoderStreamState (~135 LOC), the inner stream state machine that drives a ParquetPushDecoder. Distinct from the outer (open-file) state machine they previously shared a file with.

2. file_format.rs (~2,000 LOC) split

• New sink.rs — ParquetSink + the parallel-write machinery (column_serializer_task, spawn_column_parallel_row_group_writer, output_single_parquet_file_parallelized, concatenate_parallel_row_groups, etc.). The historical file_format::ParquetSink path is preserved via pub use.
• New schema_coercion.rs — Arrow-schema coercion utilities (apply_file_schema_type_coercions, coerce_int96_to_resolution, coerce_file_schema_to_view_type, coerce_file_schema_to_string_type, transform_schema_to_view, transform_binary_to_string, field_with_new_type) and their tests. Re-exported at the crate root for backward compat.

3. row_group_filter.rs (~1,900 LOC)

• New bloom_filter.rs — BloomFilterStatistics (the loaded SBBF data + its PruningStatistics adapter). Separates "data we loaded from the file" from "the access-plan filter that consumes it."

Net effect

File	Before	After	Δ
opener.rs (now opener/mod.rs)	2,717	2,469	-9%
file_format.rs	2,038	664	-67%
row_group_filter.rs	1,929	1,768	-8%

Are these changes tested?

Yes:

• All existing datafusion-datasource-parquet unit tests pass (the coerce_int96_to_resolution_* tests moved with the function to schema_coercion.rs).
• cargo fmt --all, ./dev/rust_lint.sh, cargo clippy -p datafusion-datasource-parquet --all-targets --all-features -- -D warnings all pass.
• Downstream datafusion core builds clean.
• datafusion-proto (which imports ParquetSink via the historical file_format::ParquetSink path) builds clean thanks to the pub use re-export.

Are there any user-facing changes?

No. Public API is unchanged — every previously-public item is still reachable at the same crate-root path. The only difference is the file organization inside the crate.

[alamb]

We could potentially split this into module moves too

[adriangb]

I'm splitting this into smaller PRs.

[adriangb]

Closing in favor of splitting this into three smaller, independently-reviewable PRs (per the request in the discussion to keep each move reviewable on its own):

• refactor(parquet-datasource): split opener.rs into an opener/ module #22346 — split opener.rs into an opener/ module (early_stop.rs, encryption.rs)
• refactor(parquet-datasource): split sink and schema_coercion out of file_format.rs #22347 — split sink.rs and schema_coercion.rs out of file_format.rs
• refactor(parquet-datasource): split bloom_filter out of row_group_filter.rs #22348 — split bloom_filter.rs out of row_group_filter.rs

Each is pure code motion, rebased on current main, and builds/tests green on its own.

One change from this PR: the opener/push_decoder_stream.rs extraction is dropped — #22289 has since extracted PushDecoderStreamState into push_decoder.rs, making that move obsolete.

[alamb]

Closing in favor of splitting this into three smaller, independently-reviewable PRs (per the request in the discussion to keep each move reviewable on its own):

And by the time I found this they are all merged 🎉