feat: Make parquet FileMetadata prunable for IR-plan dispatch

[azimafroozeh]

Make parquet FileMetadata prunable to ship in IR plans, so distributed workers can avoid re-fetching footers from cloud storage and re-decoding thrift per query. Pruning drops non-projected columns, stats from non-predicate columns, and row-group fields with no read-path consumers; an env-gated optimizer pass (POLARS_PRUNE_PARQUET_METADATA=1, default off) calls pruned() at the end of optimize(). Single-node users see no behavior change; the table below shows what would ship per query when the gate is enabled.

TPCH SF=10 wire bytes per query

table	Q1	Q3	Q5	Q6	Q19
lineitem	80.1 KB (11.8×)	52.3 KB (18.0×)	44.2 KB (21.4×)	78.1 KB (12.1×)	104.5 KB (9.0×)
orders		12.5 KB (12.2×)	10.6 KB (14.4×)
customer		1.0 KB (16.5×)	664 B (25.8×)
part					2.7 KB (7.9×)
supplier			90 B (13.0×)
nation			93 B (7.8×)
region			86 B (7.2×)
TOTAL ship	80.1 KB (11.8×)	65.9 KB (16.9×)	55.7 KB (20.0×)	78.1 KB (12.1×)	107.2 KB (9.0×)
vs re-fetch	943.7 KB	1113.7 KB	1116.2 KB	943.7 KB	964.9 KB

[codecov]

Codecov Report

❌ Patch coverage is 8.24742% with 267 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.30%. Comparing base (4808230) to head (d74ca87).

Files with missing lines	Patch %	Lines
...arquet/src/parquet/metadata/file_metadata_serde.rs	0.00%	151 Missing ⚠️
...lars-parquet/src/parquet/metadata/file_metadata.rs	0.00%	55 Missing ⚠️
...plan/src/plans/optimizer/parquet_metadata_prune.rs	22.22%	28 Missing ⚠️
crates/polars-python/src/functions/io.rs	0.00%	18 Missing ⚠️
...-parquet/src/parquet/metadata/schema_descriptor.rs	0.00%	11 Missing ⚠️
...quet/src/parquet/metadata/column_chunk_metadata.rs	25.00%	3 Missing ⚠️
crates/polars-parquet/src/arrow/read/statistics.rs	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #27535      +/-   ##
==========================================
- Coverage   81.42%   81.30%   -0.12%     
==========================================
  Files        1837     1839       +2     
  Lines      255165   255449     +284     
  Branches     3179     3179              
==========================================
- Hits       207759   207685      -74     
- Misses      46582    46940     +358     
  Partials      824      824              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[nameexhaustion]

blocking:

Could we make a lookup for 1 side (either the kept names or the parquet schema names)? We've had cases in the past where users had >=10,000 columns and saw quadratic slowdowns on lookups.
If we do kept names, I'm thinking maybe HashMap<PlSmallStr, bool>, where the bool indicates true if we want to keep statistics and false otherwise.

[nameexhaustion]

suggestion:

Could we (if it's feasible) serialize to / return a JSON string, then add a test in Python that inspects e.g. unused columns are pruned, and statistics are only kept for predicate cols?