trie/bintrie: flag-byte serialization for stem padding (follow-up to #34055)

[CPerezz]

In PR #34055 review comment 3101638279, @gballet suggested on `trie/bintrie/store_commit.go:157` (stem-node serialization):

might be worth putting some flag in the first byte instead of storing 0s. That should save a lot of disk, even though the compression code should also take care of that I guess.

Tracked as a follow-up because it's a wire-format change that shouldn't be conflated with the current arena / refactor PR:

• Changing the stem-encoding byte layout is backward-incompatible with existing bintrie blobs on disk.
• Needs a versioning story (explicit format flag or on-open migration).
• Orthogonal to the `NodeStore` unexport / representation-revert work in trie/bintrie: replace BinaryNode interface with GC-free NodeRef arena #34055.

Scope here:

1. Define the flag-byte encoding (which bits mean what, how empty slots are distinguished from present-but-zero).
2. Bump a format version so existing DBs can be detected at open and either migrated or rejected cleanly.
3. Measure the actual disk-size delta vs. the compressed baseline — gballet's own hedge ("the compression code should also take care of that I guess") means the win may be marginal post-compression.
4. Implement + benchmark + land.

Blocker/depends-on: #34055 merging first (or at least the `store_commit.go` refactor in it stabilising) so we're not fighting conflicts.

cc @gballet @CPerezz