Do not discard amended format when f-interpolator warns

[som-snytt]

Also use the same error of kind Interpolation for warnings in s.

There are assorted errors that f emits, and a handful of warnings (about format string edge cases).

Previously, f would discard the amended "parts" and adapted "args" on any diagnostic; the transform would just call format directly, so the output looks normal. But the toString warning revealed that it loses the "injected" %s specifier; that would be witnessed only if one of the rare "edge case" warnings happened to be emitted. Now it does not discard on warnings, only on hard errors.

Fixes #23693

[som-snytt]

Note to self: it would have been easier to see what was going on if FormatInterpolatorTransform were inlined in StringInterpolatorOpt#transformF. It's confusing that Transform.checked dispatches to FormatChecker#checked. I see the components were confusing before the refactor or forward port, but it could be simpler; that is, extra layers harms local reasoning. It's also nice when names are different.

Edit: went ahead and inlined that code. The other suboptimal thing is "string is empty" means "error".

[sjrd]

Just one comment on a comment, but otherwise LGTM

[sjrd]

I guess "and runtime result" is inaccurate here, isn't it?

[som-snytt]

Because I had to split the test between warn and run, for the rest of time, or until the test rig is improved, these tests must remain linked like "Rey and Ben" in Star Wars.

Edit: Rey. I must have been thinking "Jen and Ben".

[som-snytt]

Tweaked the test and rebased.

I was not ambitious to make a DottyTest that would take assertions about diagnostics and also assertions about runtime; it would be possible to assert something about the trees after the interpolator phase.