Introduce the Env::nested_read_txn from an RwTxn

[Kerollmops]

This PR is a first attempt at patching LMDB to generate multiple RoTxn from an RwTxn and, therefore, be able to read the content of an uncommitted RwTxn in parallel. We can run many algorithms in Meilisearch to process the written content and generate new data structures before committing the transaction.

// opening a write transaction
let mut wtxn = env.write_txn()?;

// [use the RwTxn to write into the env]

// opening multiple read-only transactions without committing beforehand
let rtxns = (0..1000).map(|_| env.nested_read_txn(&wtxn)).collect::<heed::Result<Vec<_>>>()?;

// [use the RoTxns and move them onto different threads]

This first attempt uses a broken fork of LMDB that ignores the number of nested write transactions associated with the parent one. Therefore, it lets heed create multiple nested write transactions from the parent one (dangerous). We store those write transactions in RoTxns shells to disallow writing anything with them at the compilation level.

However, while thinking about this, it is probably still possible to open databases (already existing) and commit them from multiple threads concurrently. It should be either disallowed or will, anyway, be fixed by the correct LMDB patch later.

To fix the Windows CI we will have to use the experimental stdatomic headers. Actually, no. We had to enable the /std:c11 and /experimental:c11atomics features.

Note that the CI is broken, but it works locally, so it's maybe a submodule issue (it's always a submodule issue) 🤔 It was...

To do

• Speed up the new nested-rtxns example in the CI (build it in release)