Preserve Parquet reader progress and add opt-in bounded range cache

[Vedin]

Summary

• Keep cold Parquet metadata readers independently pollable: a reader whose
  parent has paused it must not hold up another input reading the same file.
  Contention now performs an independent read rather than waiting on that
  reader's gate. Add a bypass metric and a deterministic paused-loader regression.
• Add an optional byte-bounded immutable Parquet range cache, disabled by default
  (ICEBERG_PARQUET_DATA_CACHE_MB=0). Keys include store-qualified file identity,
  file size and exact range. Misses preserve batched I/O; retained ranges own
  compact buffers, with entry/key overhead included in the bound.
• Document capacity, memory headroom, immutable-file assumptions, cancellation
  and the intentional duplicate-cold-I/O tradeoff. Preserve current row-lineage
  code and existing metadata cache defaults.

Verification

• Fresh local cargo fmt --all -- --check: PASS.
• Fresh local library Clippy with -D warnings: PASS on DataFusion 55.1.0 with
  the Rustice Arrow integration. Cargo metadata verifies a single 31-package
  DF55.1 closure (no mixed DF55.0/55.1 packages).
• All 28 actual-module cache unit tests: PASS, zero ignored. The local test
  wrapper compiles the two real source files, while the library Clippy run checks
  their real reader-factory integration. The wrapper avoids unrelated bundled
  DuckDB/Python test dependencies; this is not a claim that full repository CI
  has run locally. The PR's full CI remains required.
• Clippy of the actual cache modules including their tests, with -D warnings:
  PASS. The first CI run caught two single_range_in_vec_init lints in test
  setup; ranges are now constructed explicitly without changing assertions or
  allowing the lint. Runtime code is unchanged by this follow-up.
• Source hashes and byte-for-byte publication/snapshot comparisons checked before
  and after the local gate (iceberg55-r9, Rust 1.95.0).

The selected implementation was exercised during the earlier E01–E21 campaign
on TPC-H SF100 and ClickBench100m, locally and on SPCS. Those are historical
end-to-end results, not an isolated cache speedup and not qualification of the
current-main integration. A 1 GiB range-cache trial regressed its screen, so the
range cache remains opt-in. Fresh integrated benchmarks are a separate Rustice
merge gate. No raw research logs or result workbooks are included in this PR.

Limits

The footer cache trades possible duplicate cold reads for progress; retained
metadata stays bounded, but in-flight/query-owned memory needs separate headroom.
The range cache does not merge overlapping ranges or deduplicate concurrent
misses. Same-URI/same-size mutable files are not supported by its identity model.
No query-result caching, new runtime dependencies, detached background work, or
test skips are introduced.