Cardano Updates

We use `serialHasBlockIO` in the `sim` sub-library and so we have to expose it
somewhere, but we'd actually prefer it if it was not part of the public API.
Exposing it only from an internal module makes it clear that it should ideally
not be used by users.

We remove documentation about implementation specifics from the `HasBlockIO`
type, but we'll include it in the `IO` module in one of the next commits.

It's specific to the `IO` implementation

As I added this, I noticed `OffsetKey` is included in `CursorTrace`, but this is
an internal type, so I changed it to `Maybe RawBytes` instead.

The employed trick is to use no-break spaces, which is a hack but works.

Now that `topBits64` is fully safe, remove the 8-byte key constraint when using
the compact index. Instead, the config option for the index type includes a hint
not to use the compact index if their keys are too small, because that will lead
to bad performance.

When the size of the raw bytes is less than 8, then a fallback method is used
that is likely slower. When the size of the raw bytes is 8 or higher, then an
additional integer comparison is made, but this has very little impact on
performance, which I verified using micro-benchmarks. The upside to making the
function safe for any input raw bytes is that the API and test generators become
simpler, because there is one fewer constraint to satisfy: the minimum size of 8
bytes for serialised keys.

Now that `topBits64` is fully safe, remove the 8-byte key constraint when using
the compact index. Instead, the config option for the index type includes a hint
not to use the compact index if their keys are too small, because that will lead
to bad performance.

A small bug related to the merge schedule and union levels was unearted after
recent changes to key generators for the state machine tests.

```
❯ cabal run lsm-tree-test -- --quickcheck-replay="(SMGen 13447116701882578385 16048829213438376903,35)"   -p '$NF=="propLockstep_RealImpl_MockFS_IO"'
lsm-tree
  Test.Database.LSMTree.StateMachine
    propLockstep_RealImpl_MockFS_IO: FAIL (28.54s)
      *** Failed! Assertion failed (after 1 test and 196 shrinks):

      ...

1 out of 1 tests failed (28.54s)
```

The bug was that the last level was always removing `Delete` entries even if
there was union level. This is now fixed.

When the size of the raw bytes is less than 8, then a fallback method is used
that is likely slower. When the size of the raw bytes is 8 or higher, then an
additional integer comparison is made, but this has very little impact on
performance, which I verified using micro-benchmarks. The upside to making the
function safe for any input raw bytes is that the API and test generators become
simpler, because there is one fewer constraint to satisfy: the minimum size of 8
bytes for serialised keys.

Don't forget to release `HasBlockIO` in the public API

GHA: fix caching in `lint-generate-readme` job

That is, `lsm-tree-bench-wp8` becomes `utxo-bench` and `rocksdb-bench-wp8`
becomes `utxo-rocksdb-bench`. We put `bench` as the last word because it is
consistent with naming of other components in the `lsm-tree` library, e.g.,
`lsm-tree-test`, `lsm-tree-micro-bench`, etc.