fix: avoid char-boundary panic in NBReader::try_read

[SAY-5]

The internal buffer stores raw input bytes as latin-1 chars, so a byte >= 0x80 becomes a two-byte char and drain(..1) could land off a char boundary and panic. Drain the first char by its UTF-8 length instead.

Closes #32

[SAY-5]

Split into a test commit that shows the panic on multi-byte input followed by the fix, per the contrib guide.

[SAY-5]

Restructured per the c-test guide: the test commit now documents the panic with should_panic, and the fix commit drops should_panic and asserts the decoded chars. Pushed.

[SAY-5]

Force-pushed without the thread::sleep, both tests now use a deadline-bounded poll on try_read so they finish as soon as the reader thread has pumped the two bytes. test suite still passes locally.

[SAY-5]

Switched to a synchronous fixture that fills the buffer directly, no sleeps or polling.

[epage]

Generally, I would be against breaking abstractions in tests but this might be the least bad option, so moving forward with it. We can always reevaluate in the future.