storage: implement StateRead::nonverifiable_range

[erwanor]

This PR adds support for range queries over the nonverifiable storage. It does so by:

• extending the StateRead trait with a new method:

    /// Retrieve all values for keys in a range from the non-verifiable key-value store, as raw bytes.
    /// This method does not support inclusive ranges, and will return an error if passed one.
    ///
    /// Users should generally prefer to use wrapper methods in an extension trait.
    fn nonverifiable_range_raw(
        &self,
        prefix: Option<&[u8]>,
        range: impl RangeBounds<Vec<u8>>,
    ) -> Result<Self::NonconsensusRangeRawStream>;

• use a rocksdb range iterator to create an underlying stream of keys and values
• implements a StateDelta-level stream that interleaves caching layers and underlying storage.

Example API:

The first intended use of range queries is to fetch a range of compact_blocks from an arbitrary height:

let start_height = compact_block::height(20); // "000000000000000020"
// Return a stream of compact blocks starting at height 20 to the latest one.
let mut compact_block_stream = state.nonverifiable_range_query(Some(compact_block::prefix()), start_height..);

Alternative APIs considered:

        fn nonverifiable_range_raw<T, U>(
        &self,
        prefix: Option<&[u8]>,
        range: T,
    ) -> anyhow::Result<Self::NonconsensusRangeRawStream>
    where
        T: std::ops::RangeBounds<U>,
        U: Into<Vec<u8>>;

and

    fn nonverifiable_range_raw<T, U>(
        &self,
        prefix: Option<&[u8]>,
        range: T,
    ) -> anyhow::Result<Self::NonconsensusRangeRawStream>
    where
        T: std::ops::RangeBounds<U>,
        U: AsRef<[u8]>;

either are fine choices, but my preference goes to the simpler signature.

[hdevalence]

Is there any reason not to merge this for the current testnet?

[erwanor]

Not really, I left it open to invite reviews but merging now