Preserve evolved manifest bounds in DF55 fork

[osipovartem]

Summary

• Port the consumer-boundary design from Decode manifest bounds after schema promotion JanKaul/iceberg-rust#410 to the DF55 fork: decode manifest bounds using the embedded schema and promote them only for DataFusion statistics and pruning.
• Accept old 4-byte numeric/date bounds under promoted types without automatically rewriting those bounds to current widths.
• Preserve the fork's cached historical partition-type candidate decoder. When old source fields are no longer available, retain the manifest and discard only optional summaries. Mixed partition specs produce unknown pruning stats instead of false exclusions; same-spec old bounds remain prunable.
• Propagate manifest read/serialization errors in the rewrite path instead of silently skipping entries.

The replacement avoids the former selected-schema threading, public ManifestFieldTypeIndex, and rewrite-only reader mode. A follow-up benchmark should quantify the conservative manifest-pruning cost on tables retaining many old partition specs; correctness takes precedence over guessing whether fields with different specs are compatible.

Local validation

• cargo +1.95.0 test -q -p iceberg-rust-spec (177 active passed)
• cargo +1.95.0 test -q -p iceberg-rust --lib (172 active passed)
• cargo +1.95.0 test -q -p datafusion_iceberg --lib (89 passed)
• cargo +1.95.0 test -q -p iceberg-rust --test overwrite_test (3 passed)
• cargo +1.95.0 clippy -p iceberg-rust-spec -p iceberg-rust -p datafusion_iceberg --all-targets -- -D warnings
• cargo +1.95.0 fmt --all -- --check
• Independent read-only review approved after addressing the retained-manifest/list decode case.

This PR does not update Rustice's dependency pin.

[osipovartem]

Local pruning microbenchmark before merge: 10,000 current-schema data-file bounds, 30 min_values passes per trial, five trials, identical debug test profile. Median baseline (c32b9e0): 74.37 ms; median PR head (bef12e3): 73.62 ms. No measurable regression on the common current-schema path. Historical incompatible partition specs intentionally produce unknown pruning stats to avoid false negatives; such tables may read more manifest metadata until per-spec pruning is added. Temporary benchmark code was removed; the PR worktree is clean.