Fix TrieDBMut panic when removing a strict-prefix key

[dimartiro]

Summary

Removing a key that is only a strict prefix of existing keys panics in debug builds and silently corrupts the trie in release builds, under any no-extension layout (including the substrate-like hashed-value codec used in production).

Closes #217.

Reproduction

// NoExtensionLayout
let mut memdb = MemoryDB::<_, HashKey<_>, _>::default();
let mut root = Default::default();
let mut t = TrieDBMutBuilder::<NoExtensionLayout>::new(&mut memdb, &mut root).build();
t.insert(&[0xff, 0xff], &[0x0a]).unwrap();
t.insert(&[0xfe, 0xff, 0xe9], &[0x0b]).unwrap();
t.insert(&[0xff, 0xff, 0xff], &[0x0c]).unwrap();
t.remove(&[0xff]).unwrap(); // `ff` is absent — panics: assertion failed: self.len() >= i

Root cause

ff is not present but is a strict prefix of ffff / ffffff. While descending, the root NibbledBranch consumes the whole search key via key.advance(common + 1), so the recursion reaches the inner NibbledBranch (whose own partial the key never traversed) with an empty remaining key.

The (Node::NibbledBranch(n, children, val), true) arm in remove_inspector treated "empty remaining key" as "the value lives here". It removed a different (longer) key's value and called fix, whose fix_inner advances the nibble offset by the branch's own partial length on an already-exhausted key — tripping debug_assert!(self.len() >= i) in NibbleSlice::advance (panic in debug; in release the offset overflows and a valid value is dropped).

Fix

Guard the value removal on the branch's own partial being empty. A NibbledBranch's value is reachable only by a key that still contains the whole of its partial, so an empty remaining key with a non-empty partial means the key terminated early and is absent — removal must be a no-op (Action::Restore).

Testing

Two regression tests, run across all reference layouts via test_layouts! (including HashedValueNoExtThreshold, the substrate-like hashed-value codec):

• remove_prefix_key_is_noop_issue_217 — removing the absent prefix key leaves the trie root byte-identical and all existing values intact.
• remove_exact_branch_value_issue_217 — a value genuinely sitting on a partial-bearing branch is still removable (the guard must not over-skip legitimate removals).

Full trie-db-test suite passes (126/126).