Optimize read-after-read access in `Matrix` subpipelines

[msimberg]

See #898 (comment). In the worst case this may need support in async_rw_mutex in pika. Needs further investigation. Some investigation on where and if this actually could lead to a performance benefit may be useful.