fix: correct midpoint level so odd-depth uploads don't revert

[grumbach]

Summary

• Fixes WrongPoolCount reverts at the on-chain payForMerkleTree call for files producing odd-depth merkle trees (chunk counts in the bands 5-8, 17-32, 65-128). Production observed it as WrongPoolCount(16, 8) for files in the 65-128 chunk band.
• Root cause: MerklePaymentLib.expectedRewardPools(depth) = 2^ceil(depth/2) (Solidity), but the client computed midpoints at level ceil(depth/2) from the leaves. ant_merkle::get_nodes_at_level numbers levels from the leaves up (L=0 leaves, L=depth root), so picking L = ceil(d/2) returns 2^floor(d/2) nodes — half the count for odd depths.
• Fix: set midpoint_level = floor(depth/2), which yields 2^(d - floor(d/2)) = 2^ceil(d/2) nodes — matches the contract for both even and odd depths. expected_reward_pools is unchanged (already 2^ceil(d/2) since fix: payment hardening -- pricing, panics, external signer #5).
• Adds five regression tests that exhaustively cover every supported leaf count, including the production failure mode and the full verifier path.

Why the previous attempt missed this

PR #5's commit f4cdd45 ("align pool count formula with Solidity contract (floor->ceil)") flipped both expected_reward_pools and midpoint_level to ceil in lockstep. That fixes expected_reward_pools (correct) but leaves the bug in midpoint_level because ant_merkle levels are counted from the leaves — flipping floor to ceil there moves the cut closer to the root and halves the pool count for odd depths. The doc comment now spells this out so the next reader doesn't repeat the mistake.

Test plan

• cargo test --lib — 16 passed (5 new merkle regression tests + 11 pre-existing)
• cargo clippy --all-targets -- -D warnings clean
• cargo fmt clean
• Verify upload of a file in each failing chunk band (e.g. 25 MiB, 100 MiB, 400 MiB) succeeds against the live testnet once this lands and clients pick it up

Risk

• The fix is one line (depth.div_ceil(2) -> depth / 2) in a non-public function. No wire-format change, no contract change.
• Even-depth trees are unaffected: floor(d/2) == ceil(d/2) for even d.
• midpoint_proof_depth is unchanged and remains ceil(d/2) — that is the path length from level floor(d/2) to the root, which is d - floor(d/2) = ceil(d/2). The every_midpoint_proof_verifies_for_all_depths and verify_merkle_proof_accepts_all_depths tests confirm the proof shape stays consistent post-fix.