Some call places do not need calls to 'node_sanitize' anymore, therefore
removing the call entirely.

Other still use it for convenience, but that doesn't seem the right way
to do it for consistency. Those places have been replaced by calls to
'Node.strip_node_info()'.

Remove completely '_yaml.dump()' and replace all notions and call by
'roundtrip_dump'

Now that both are equivalent, we can skip the sanitization part before
the yaml call.

This removes the need of calling _yaml.dump(), as roundtrip_dump
is now equivalent.

This removes the need of manually stringifying the values in
roundtrip_dump by registering special Ruamel representer for each
value we might expect.

Since ujson already sorts the keys, we just need to be able to jsonify
Nodes sensibly. This is done by implementing __json__ on the 'Node' base
class.

This does not break the cache keys.

- node_find_target with 'key' is only used once in the codebase.
  We can remove and simplify this function
- Allow 'MappingNode.get_node()' to be called without any
  'expected_types'

- Implement __setitem__ on 'MappingNode'
- Implement __setitem__ on 'SequenceNode'
- Adapt all call sites to use the new calling way.

One difference is that 'MappingNode.items()' does not strip the
provenance from scalars and lists, which ends up not affecting the
code much.