fix: propagate validity in RunEnd and Sparse lazy decoders (#225, #226)

[dfa1]

Fixes the two round-4 silent-null-loss bugs (#205 triage), same class as the fixed #210: the RunEnd and Sparse lazy decoders unwrapped their children's MaskedArray to .inner() and dropped validity, so null rows decoded as plausible values.

Semantics translated directly from the Rust ValidityVTable impls (encodings/runend/src/array.rs, encodings/sparse/src/lib.rs):

• RunEnd (RunEndEncodingDecoder drops nullable run-values validity — silent null loss #225): a row's validity is the validity of the run it falls in → row validity is a RunEnd bool over the same ends whose per-run value is the run-value's validity bit. Represented lazily (LazyRunEndBoolArray), O(runs) — essential since n can be huge (uci-online-retail 499,712 rows); a per-row bitmap would defeat the lazy path.
• Sparse (SparseEncodingDecoder drops nullability — null fill coerced to 0, nullable patch values stripped #226): a position is valid iff (it's a patch AND that patch is valid) OR (it's unpatched AND the fill is non-null). Represented lazily (LazySparseBoolArray), O(patches). Two facets fixed: the null fill value (previously scalarToLong coerced null→0, so unpatched positions rendered 0.0), now detected from the ProtoScalarValue itself (a real 0 fill sets a typed field; null sets null_value/none); and the stripped nullable patch_values mask.

Validation: new RunEndEncodingDecoderTest (3) + SparseEncodingDecoderTest (4) covering null-run/null-fill/null-patch and the no-regression non-nullable path; 2 SparseEncodingEncoderTest round-trips updated (they asserted the old null-fill-as-zero behavior). Real corpus: uci-online-retail (4.3M cells) and world-energy-consumption (3.0M cells) both match the Parquet oracle exactly. Matrix: both gap:225/gap:226 → ok (0 gaps). Full ./mvnw verify green.

Known follow-up (out of scope, filed separately): VarBin/Utf8 sparse (decodeVarBin) still drops null-fill validity — same class, no corpus column hits it yet.

Closes #225
Closes #226

🤖 Generated with Claude Code

[dfa1]

Reviewed (vortex-reviewer, built at ref + ran suites): APPROVE, no correctness/memory blocker. Validity propagation verified faithful to the Rust ValidityVTable contract across all edges — all four Sparse fill×patch combinations, RunEnd first/last/length-1 runs + offset, and isNullScalar confirmed complete over all 11 ProtoScalarValue components (integer 0 / 0.0 / false / "" correctly non-null — no false-negative). Lazy O(runs)/O(patches), no-regression branches return bare arrays. Three nits, non-blocking: (1) isNullScalar could drift if the proto gains a field — folding a keep-in-sync guard into the stacked #232 work which touches the same helper; (2) no gate-running interop fixture (only the weekly corpus flip) — filing as a follow-up; (3) harmless discard-after-decode in the RunEnd unsupported-dtype path.