rtk diff: line-ending and trailing-newline differences go unreported when a line change coexists

Split out of #3268, where @KuSh found it during review. It is not a regression: develop gates on the same condition.

InvisibleDifference is reached only when the line-level diff is empty, so a CRLF or trailing-newline difference that coexists with a content change is never mentioned.

file1 = "alpha\nbeta\n"   file2 = "alpha\r\ngamma"

rtk:  ~   2 beta → gamma
GNU:  1,2c1,2   (line 1 differs too)  +  "\ No newline at end of file"

A reader concludes that line 1 is byte-identical and that both files end in a newline. Neither is true. The exit code is 1 either way, so this is narrower than the false-identity bug #3268 closes, but it is the same class: a difference str::lines() normalises away is not reported.

tests/diff_byte_accuracy_test.rs covers the pure-invisible case only.

Sketch: run the CRLF-position and trailing-newline checks in describe_invisible_difference on every comparison rather than only when the hunks are empty, and when the hunks are non-empty append the result as one indented note under the listing. The note is a few dozen bytes and fires only when such a difference exists, so never_worse should rarely see it, but that needs measuring on the 120-diff corpus before it ships.

Start with describe_invisible_difference and tests/diff_byte_accuracy_test.rs, which currently covers only the pure-invisible case. Trace how comparisons with non-empty hunks are formatted, then add coverage for the CRLF-position and trailing-newline cases described here. Check the never_worse result against the 120-diff corpus; done means the coexistence differences are reported without regressing that corpus.