use of boost::reversed_iterator within boost::reversed_range adaptor

[smitsyn]

We use custom span class template, and it works incorrectly with boost::adapters::reverse. Link to reproduce:

https://godbolt.org/z/1Yxvqvj4r

Executable compiled with clang prints incorrect value, with gcc - exits with nonzero code, and msvc compiler emits a warning. I believe godbolt doesn't support sanitizers ATM. Non-const span is provided and works fine.

I believe the problem is that iterator_range::operator[] returns by reference and uses boost::reverse_iterator::operator[], but the latter reserved a right to return by value. According to docs ( https://www.boost.org/doc/libs/1_79_0/libs/iterator/doc/reverse_iterator.html ) it is "Random Access Iterator", and according to other docs ( https://www.boost.org/doc/libs/1_79_0/libs/iterator/doc/new-iter-concepts.html#design ), "operator[] is only required to return something convertible to the value_type (for a Readable Iterator)", which reverse_iterator does ( https://www.boost.org/doc/libs/1_79_0/libs/iterator/doc/new-iter-concepts.html#random-access-traversal-iterators-lib-random-access-traversal-iterators ): "a[n] has return type convertible to T" (for random access iterator as defined in now quite old "New Iterator Concepts").

Note that some std::reverse_iterator implementations return by reference in const T case too and could work with boost::iterator_range.

[smitsyn]

When #122 is merged, this could fail with std::span too.