Drop odd-length flat dead byte; make --analyze score consistent

[behdad]

Two low-severity cleanups from the earlier soundness review.

1. Odd-length flat "dead byte"

InnerLayer.split() padded self.data in place to make pairing work, but
the flat (unsplit) solution emits self.data directly — so odd-length flat
tables carried one unreachable trailing padding byte, and the reported
.cost (computed from the un-padded length) was one element short of the
emission.

Fix: pad a local copy instead. self.data stays the original array, so the
flat solution emits exactly len(data) elements. Split solutions are unchanged
(pairing still uses the optimally-padded copy), so solution costs — and
pick_solution's choices — are identical; only the flat array for odd-length
tables loses its dead byte (one byte smaller, never larger). No runtime ops are
added to the generated code; generation cost is a negligible O(n) copy for odd
layers only.

2. --analyze score display

The displayed Score used floor(log2) (fullCost.bit_length() - 1) while
pick_solution uses exact log2, so the highlighted "Best solution" could
disagree with the minimum-score row (e.g. two solutions whose fullCosts share a
floor(log2) bucket). Use exact log2 and show two decimals so the ranking is
visibly consistent.

Verification

• pytest: 236 passing (adds tests: flat odd-length emits no padding byte,
  split() doesn't mutate self.data, flat odd-length round-trips in C+Rust,
  and --analyze best-row == min-score-row).
• A C+Rust round-trip fuzz over all Pareto solutions: zero failures.
• Regenerating HarfBuzz's Unicode tables (make -f update-unicode-tables.make clean all) yields byte-identical output in ~10s (the flat fix is a no-op
  for HarfBuzz's multi-level tables).

🤖 Generated with Claude Code