BREAKING CHANGE: v0.0.6 -> --randomize/-R impl in chaintools split

[alejandrogzi]

Added

• --randomize/-R (with optional --seed <u64>) for chaintools split. When
  enabled, chains are distributed across output files in random order instead of input
  order, so when the input is sorted by id/score the largest chains spread evenly across
  files instead of all landing in the first one.
• Hand-rolled RNG, no new dependency — SplitMix64 PRNG + Lemire-bounded
  next_bounded + Fisher–Yates shuffle_indices, plus a time-based default_seed()
  (reusing the already-imported SystemTime). Seed is logged at info level so every
  run is reproducible.
• OutputPlan.byte_range: Range → byte_ranges: Vec<Range>. The non-randomized path
  is untouched and just as fast: it still produces a single contiguous zero-copy range.
  The randomized path shuffles a chain-index permutation, partitions it with the existing
  --files/--chunks math, then merge_chain_ranges maps each file's chains to a
  minimal set of byte ranges (sorted + contiguous-coalesced to minimize writes).
• Robustness fix — merge_chain_ranges makes chain 0 own byte 0, so the per-chain
  ranges form a gap-free partition of the input. This preserves any preamble before the
  first chain header (which the naive per-chain mapping would have dropped) and guarantees
  no bytes are lost or duplicated.
• Writer — write_output_slice now concatenates multiple ranges into the same
  BufWriter/GzEncoder, preserving zero-copy slicing and gzip support.

Changed

• CLI flags in src/cli/split.rs — added -R/--randomize (bool) and --seed (optional
  u64, with clap requires = "randomize" so --seed alone is an error).

Documentation

• Both flags documented in assets/tools/split.md.

Notes

• 82 binary tests + 67 lib/integration tests pass, including 7 new ones (permutation
  validity, determinism, range merging, preamble preservation, chain preservation,
  reproducibility, and seed-without-randomize rejection). gzip-feature tests pass too.
• End-to-end: confirmed reproducibility (--seed 42 byte-identical across runs),
  fresh-seed logging, all chains present exactly once, and visible redistribution of
  chains across files.