feat(miner): add validate_deadlines flag to recheck new best deadlines

[JohnnyFFM]

Summary

Adds an opt-in validate_deadlines config flag (default false) that independently re-derives every new best deadline before the miner trusts it — a best-effort safeguard for users with failing hard drives.

Problem

A failing disk can read a scoop back corrupted, which then hashes to an essentially random — and sometimes lower (better) — deadline. Today that bogus deadline is:

1. submitted to the chain, where it fails verification (wasted submission); and worse,
2. recorded as the account's best_quality, so a genuine, slightly-higher deadline found later (in another warp/file) is filtered out by the adjusted_quality < best_quality gate. The bad solution overshadows a valid one locally.

Fix

When validate_deadlines is enabled, the nonce-submission task — the single place that tracks and updates best — re-derives the canonical scoop from (account_id, seed, nonce, scoop) through the scalar (non-SIMD) code path, recomputes the quality, and compares it against the on-disk value. On mismatch (or any decode/regeneration error) the candidate is rejected: best is not updated and nothing is submitted, so a genuine solution can still win.

The scalar and SIMD quality computations are bit-identical (already asserted by existing tests), so a healthy scoop always reproduces its quality exactly; only corrupt data mismatches. Re-derivation runs only on the rare new-best event over a single small scoop — a fair tradeoff for the protection it provides.

Changes

• pocx_hashlib: refactor generate_scoop to share its loop, add scalar generate_scoop_scalar, calculate_quality_raw_scalar, and calculate_quality_from_height_scalar (built on generate_nonces_32 + find_best_quality_32).
• pocx_miner: add validate_deadlines (default false) to Cfg, CfgBuilder, and miner_config.yaml.
• pocx_miner: add the revalidate_deadline gate in create_nonce_submission_tasks.
• tests: scalar/SIMD quality equivalence and corruption detection.

Known limitation

find_best_quality returns only the single minimum per warp buffer. If a corrupt scoop and a genuine better scoop live in the same warp buffer, rejecting the corrupt one does not recover the genuine one (it was never surfaced). This PR fixes the cross-warp/cross-file overshadowing, which is the main harm; intra-warp masking would require re-scanning inside the hasher and is out of scope.

Verification

• cargo test -p pocx_hashlib — 53 passed (incl. new test_scalar_quality_matches_simd, test_validate_deadline_detects_corruption).
• cargo test -p pocx_miner — 29 passed.
• cargo fmt --all --check — clean.
• cargo clippy -p pocx_hashlib -p pocx_miner --all-targets -- -D warnings — clean.