feat(types): rooftop solar forecast metric, 30m interval #675

[nc9]

client side of opennem/opennem#675 (rooftop solar forecast), plus two response shape fixes.

forecast:

• MarketMetric.SOLAR_ROOFTOP_FORECAST
• 30m on DataInterval and VALID_INTERVALS. NetworkTimeSeries.interval is the literal so a 30m response failed validation before this
• optional forecast_run_time on NetworkTimeSeries, None for history loaded before oct 2026
• examples/rooftop_forecast.py, base url via OPENELECTRICITY_API_URL

response shape (live api sends date_start/date_end and columns.region):

• NetworkTimeSeries.date_start/date_end and TimeSeriesColumns.region (+ status). start/end/network_region stay as deprecated aliases with the same values, kept in step on validate, assignment, copy, construct and unpickle, and they warn on read
• to_records()/to_pandas()/to_polars() no longer shift intervals +10h and include region (and unit_code for facilities). default row shape is unchanged, one row per value. new keyword-only merge_metrics=True gives one row per interval + grouping with a column per metric
• to_polars() scans every row for the schema, it dropped any metric whose rows started after row 100
• old-shape responses (start/end, network_region, utc timestamps) still parse and give the same record columns and times as before

breaking for data frame users: interval is no longer +10h and records gain region/status/unit_code. changelog has a breaking changes section with a before/after row, readme has an upgrading to 0.12 note.

readme, changelog, bump to 0.12.0. don't publish until the prod release.

pytest (56 incl. new compat + regression tests, also under -W error::DeprecationWarning), ruff, basedpyright (no new errors) pass. smoke tested against dev 4.5.17.dev2, all pass.