Expose census appends, match_type, raw payloads and rate limits

[MiniCodeMonkey]

Closes four gaps in the client's public surface. These surfaced while reviewing a contractor-built Dagster skill that had to fall back to the third-party pygeocodio package: every one of these was a reason it could not use the official client cleanly.

Ready for review. No version bump is included, so a release is a separate call once this is approved.

Changes

Census appends had no public accessor. Requesting fields=["census2023"] stored the data only on the private fields._census dict, so consumers had to write result.fields._census["census2023"]["full_fips"] to reach the single most useful append field. Adds fields.census, fields.get_census(year), fields.census_years and fields.census_data, matching how the other appends are already exposed. get_census() accepts 2023, "2023" or "census2023". The private dict and the dynamic fields.census2023 attributes are untouched.

# before
fips = result.fields._census["census2023"]["full_fips"]
# after
fips = result.fields.census.full_fips

match_type and address_lines were silently dropped. Both are returned on every v2 result but were absent from GeocodingResult, so the data was lost on parse. Both are now populated for single and batch responses. A full diff of the raw v2 JSON keys against the dataclass confirms these were the only two missing.

No way to recover the raw payload. There was no raw, to_dict(), or equivalent, and dataclasses.asdict() is lossy and leaks the private _census key. Adds raw and to_dict() to both GeocodingResponse and GeocodingResult, so pipelines can cache the untouched response and add derived columns later without re-fetching and re-paying for the lookup.

Rate limit headers were not exposed. The X-RateLimit-* headers were logged at debug level only. Adds a RateLimit model exposed as GeocodingResponse.rate_limit and Geocodio.rate_limit, parsed case-insensitively and populated before error handling so it is available on 4xx and 5xx too. Worth noting the API sends x-ratelimit-period rather than x-ratelimit-reset; both are modelled and the raw header dict is kept on RateLimit.headers.

Compatibility

Backward compatible. New dataclass fields are appended, so positional construction still works, and existing attribute access is unchanged.

Verification

• pytest tests/unit/ -- 115 passed (9 new)
• pytest tests/e2e/ -- 45 passed against the live API (4 new)
• flake8 src/ --ignore=E501 -- clean; black --check and isort --check-only clean
• mypy -- 19 errors, identical to the baseline on main, none on new lines

All four changes were also verified by hand against api.geocod.io/v2.

Reviewer notes

• fields.census returns the most recent vintage when several are present in the response. Multi-vintage callers should use get_census(year). Worth a sanity-check that this is the behaviour you want.
• raw is a real dataclass field, so dataclasses.asdict() output now includes it and roughly duplicates the payload. It is repr=False, so reprs are unaffected.
• Unmatched batch entries get raw={} rather than the {query, response} wrapper, keeping result.raw consistently "the API result object". The wrapper stays reachable via response.raw.