Refactor: extract `BaseScan` and `ManifestGroupPlanner` (#3511)

# Rationale for this change

Split out of #3364 at the reviewers' request (to @rambleraptor's and
@kevinjqliu's point that the architecture refactor and the new feature
are easier to review separately).

This PR will be followed by
https://github.com/apache/iceberg-python/pull/3512 that makes use of
this refactor to implement append scans more cleanly.

**This PR is a pure refactor with no behavioural change.** The diff
looks large, but it is almost entirely code being moved, not changed.

# Changes

- Introduce `BaseScan(ABC)` as a superclass of `TableScan`, holding the
snapshot-independent surface: `table_metadata`, `io`, `row_filter`,
`selected_fields`, `case_sensitive`, `options`, `limit`, the chaining
helpers (`select` / `filter` / `with_case_sensitive` / `update`), and
the format-converter sinks (`to_pandas` / `to_polars` / `to_duckdb` /
`to_ray`) built on an abstract `to_arrow()`.
- `TableScan` keeps everything snapshot-specific — `snapshot_id`,
`catalog`, `table_identifier`, `snapshot()`, the snapshot-aware
`projection()`, `use_ref()`, and abstract `count()` — so its existing
public surface is unchanged.
- Extract `ManifestGroupPlanner` from `DataScan` (the `_build_*`
evaluators, `_check_sequence_number`, and the manifest-entry /
file-scan-task planning) and route all of `DataScan`'s local planning —
`scan_plan_helper()` and `_plan_files_local()` — through it, so the
partition / metrics / residual pipeline lives in exactly one place.

# Back-compatibility

- `to_arrow_batch_reader` stays concrete on `DataScan` and is **not**
made abstract on `BaseScan`, so external `TableScan` subclasses that
were valid before still instantiate.
- `to_pandas` / `to_polars` become concrete defaults on `BaseScan`, and
`to_duckdb` / `to_ray` move up to `BaseScan` too. This **loosens**
`TableScan`'s abstract contract (they were abstract on `TableScan`
before) without breaking existing subclasses.
- All docstrings and explanatory comments are preserved.

# Are these changes tested?

The existing `DataScan` unit and integration tests exercise the
refactored planning path unchanged. A regression test is added for the
no-current-snapshot planning path, which now flows through
`ManifestGroupPlanner`.

# Are there any user-facing changes?

No. Some previously-`DataScan`-only converters (`to_duckdb` / `to_ray`)
are now inherited by `TableScan` as well, which is purely additive.

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Kevin Liu <kevinjqliu@users.noreply.github.com>