feat: add pandas 3 compatibility and fix CI (drop Python 3.9, pin polars <1.38)

[mattijn]

This PR adds pandas 3 compatibility and fixes a cascade of CI failures originally surfaced in #3949. The changes span three areas (1) pandas 3 support, (2) CI infrastructure, and (3) dependency pinning, which are bundled together because each group depends on the others for CI to pass.

Changes

pandas 3 compatibility

• Handle the new str dtype introduced in pandas 3
• Fix test_pandas_date_parse: loosen assertion to allow equal schemas when parse_dates=[]
• Fix sanitize round-trip test: explicitly cast df2['o'] to object after read_json
• Use float64 explicitly so DataFrames match read_json output
• Add YE / Y frequency fallback in the test suite: pandas 2 introduced YE; pandas 1.x only knows Y, pandas 3 has deprecated Y frequency

CI / build infrastructure

• Skip metadata.parquet from bit-identical comparison in build.yml, as Parquet files are not guaranteed to be reproducible across runs (SO ref)
• Drop Python 3.9 from the test matrix (EOL since end of 2025; also triggers a ValueError: Invalid frequency: YE that cannot be cleanly worked around on that version)

Dependency pinning

• Pin polars < 1.38 until the string_view regression is resolved upstream
• Tighten minimum supported pandas version for Python 3.10
• Tighten minimum supported numpy version for Python 3.10 (numpy supports 3.10 from 1.21.2)

Why one PR?

The pandas 3 support and CI fixes logically belong in separate PRs, but each group leaves CI in a failing state on its own. They are therefore submitted together so the branch can be merged, with the intention of splitting future follow-up work back into atomic PRs once the baseline is green.