fix: prefix-free node hashing - #242

…llide

get_hash concatenated keys.concat() ++ values.concat() with NO length
delimiters, so distinct content could produce identical hash input:
values ("ab","c") and ("a","bc") both concat to "abc" under the same
keys => identical root hash. For a content-addressed / merge-convergent
structure that is a correctness bug (false dedup, false merge
convergence).

Length-frame every element (u32 BE lengths — never usize, so the hash
is identical on 32- and 64-bit / wasm), delimit the key region from the
value region by count, and bind is_leaf + level so a leaf can't collide
with an internal node of identical bytes. Hash-function-agnostic: only
the ValueDigest::new INPUT changes, not the digest algorithm, and only
node identity changes — chunk boundaries (tree shape) are untouched, so
merge-canonicality and every existing test still hold.

Tests: mod prefix_free_collision_probe (fails on the old concat hash,
passes here).