Fixes #1809

b77cec05c0 fixed this problem for tagging tests, so that a test
case tagged with `[.foo]` would be parsed as tagged with `[.][foo]`.
This does the same for the test spec parsing.

Fixes #1798

Now it no longer tries to be this weird hybrid between an owning
and non-owning reference, and is only ever non-owning. This is also
reflected in its interface, for example `StringRef::isNullTerminated`
is now public, and `StringRef::c_str()` has the precondition that it
is true.

Overview of the changes:
* The `StringRef::m_data` member has been completely removed, as it
had no more uses.
* `StringRef::isSubstring()` has been made public and renamed to
`StringRef::isNullTerminated()`, so that the name reflects what the
method actually does.
* `StringRef::currentData()` has been renamed to `StringRef::data()`,
to be in line with common C++ containers and container-alikes.
* `StringRef::c_str()` will no longer silently make copies. It instead
has a precondition that `isNullTerminated()` is true.
* If the user needs a null-terminated string, they should use the
`std::string` conversion operator and call `c_str()` on the resulting
`std::string`.
* Some small optimizations in various places.
* Basic functionality is now `constexpr`.

This means that if you nest multiple random generators inside one
test case, they will not return the same sequence of numbers.

Idea taken from #1736 by Amit Herman.

Closes #1736
Closes #1734

In the future, we will also want to introduce our own
`uniform_int_distribution` and `uniform_real_distribution` to get
repeatable test runs across different platforms.

The leading/trailing whitespace is problematic because of e.g.
`WHEN` macro having preceeding whitespace for alignment, and it is
generally messy.

Credits to Phil who did lot of the original work.

Closes #1708