Pass the prefix by reference in storage key prefix iteration

[oliv3rdrt]

The prefix and key passed to prefix_iter and greater_equal_iter are only read, but they were taken by value, so callers had to hand over an owned Vec even when they already had a slice. In the locking adaptor this meant cloning the prefix twice on every call.

This changes the ReadOps methods, and the matching helpers in the database module, to take &[u8], and updates the backends (in-memory, lmdb, sqlite, failing, and the locking adaptor) and their callers.

What it removes in practice:

• the two prefix.clone() calls and the key.clone() in the locking adaptor
• Vec::new() and Data::new() allocations at call sites that only needed an empty prefix
• a couple of clone() calls in the backend test suite

One thing worth a look during review: since the crate is on edition 2024, the returned impl Iterator would otherwise capture the prefix lifetime as well as the transaction's. I used the precise capturing syntax (use<..>) on the trait methods and the impls so the iterator only borrows from the transaction, which is what the previous signature effectively guaranteed. If you would rather see it done another way I am happy to change it.

Checked with cargo check --workspace --all-targets, cargo fmt, and the storage tests (storage, storage-core, storage-inmemory, storage-lmdb, storage-sqlite) plus chainstate-storage and utxo, all passing. Note that a few of the lmdb property tests fail intermittently when run in parallel both with and without this change, which looks like the flakiness tracked in #762, they pass with --test-threads=1.

Addresses #417.

[oliv3rdrt]

@erubboli can I be assigned this?