perf(prune): backport efficient history shard pruning to v2.5.2

[0x00101010]

Why

History pruning blocks persistence while decoding retained shards that cannot contain expired data. Backports the upstream optimization onto backport/v2.5.2/main, currently tagged base-v2.5.2.7.

What

• Cherry-pick the complete merged upstream patch unchanged, preserving attribution, provenance, and regression tests.
• Stop history scans after the boundary shard and trim expired bitmap ranges in place, preserving the live sentinel and propagating iterator errors. Retention, database formats, and scheduling are unchanged.

Notes

• A separate A/B/A mainnet-snapshot trial at 200ms with a 1,339,200-block window measured 3.73 → 4.99 → 3.60 blocks/s and mean prune times of 4.20s → 0.124s → 4.33s. The backport kept pruning one old block per new block. Each run measured 120s after 30s warm-up on one host; runs advanced the same datadir, and the backport binary used local Reth dependency overrides.
• The 13,392,000-block window, longer runs, and other hosts remain untested.

[0x00101010]

Local verification: 326 tests passed across reth-db-api, reth-provider, and reth-prune; nightly workspace formatting and all-target/all-feature Clippy for those crates passed. The patch ID matches the upstream commit.

The expanded all-feature test run passed 330/331 tests. The failure, remove_block_and_execution_above_returns_persistence_frontiers, also fails on unchanged base-v2.5.2.7 with the same missing-trie-updates error. Full-workspace all-feature Clippy is blocked locally by missing LLVM 22 and m4.

The lint workflow failed at startup before creating jobs; the unchanged workflow has the same startup failure on another PR against this base. No unrelated code or workflow changes are included.

[0x00101010]

The workspace unit tests, both E2E suites, doctests, and codec checks passed. Ethereum state tests were interrupted by a runner shutdown after 34 tests passed, so the aggregate unit check failed. The previous release had the same runner-shutdown failure before a successful retry. I requested one retry of the failed jobs; no source changes are needed for this interruption.