Add CPA per-year bulk download (download_cpa_file) once OECD fixes malformed dotStat files

[jm-rivera]

Add CPA per-year bulk download (download_cpa_file) once OECD fixes the malformed dotStat files

Follow-up to #38 (CPA reader). The CPA API reader download_cpa() shipped in #38. The per-year bulk
path is blocked on an upstream OECD data bug, so download_cpa_file() is intentionally not
included yet.

Why it's blocked

The per-year "CRS CPA <year> (dotStat format)" .txt bulk files (dataflow
OECD.DCD.FSD:DSD_CPA@DF_CRS_CPA) are malformed: a large fraction of rows have more
pipe-delimited fields than the 49-column header declares, so they can't be parsed against their own
header. The same records via the SDMX API are clean.

Affected-row counts (rows whose field count ≠ 49):

Year	Rows	Ragged %	Rows with >49 non-empty fields
2023	251,992	69%	47,345
2022	233,467	48%	25,704
2021	252,976	33%	24,469
2020	188,902	35%	12,837
2015	152,954	0%	0

Tens of thousands of rows/year have more than 49 non-empty fields (structurally impossible for a
49-column row), and the API confirms the real text fields contain zero | characters — so the extra
delimiters are spurious. An extra (empty or duplicated-text) field is injected mid-row, shifting all
later columns. Not deterministically recoverable. This has been reported to the OECD.

Implementation once the upstream files are fixed

Small — clone the CRS bulk path (crs.py download_crs_file / get_year_crs_zip_id):

• Add download_cpa_file(year, save_to_path=None, *, as_iterator=False, use_raw_cache=True) to
  src/oda_reader/cpa.py, plus get_year_cpa_zip_id(year) using
  search_string=f"CRS CPA {year} (dotStat format)" and CPA_FLOW_URL (=
  BASE_DATAFLOW + "DSD_CPA@DF_CRS_CPA/").
• Export download_cpa_file from __init__.py (__all__).
• Bulk is per-year only (2010–2024); there is no all-years/full-dataset CPA bulk file, so no
  bulk_download_cpa().
• Add a unit test (mock bulk_download_parquet) + an integration test, and a README "Per-year CPA
  files" subsection.

Acceptance

• download_cpa_file(<year>) returns/saves a year of CPA data and parses cleanly (no ragged-row
  errors) once OECD has corrected the files.
• Re-verify raggedness is 0% on the corrected files before implementing.

[jm-rivera]

OECD acknowledged the problem, will issue a file correction in the next couple of weeks