fix(ansi): reset parser state after grapheme cluster in Truncate

[sueun-dev]

Fixes the Truncate part of #541.

Truncate could return a string wider than the requested length. In the Utf8State branch of truncate(), pstate was reset to GroundState only on the path that writes the cluster; the "ignoring" and "too wide" continue paths left it mid-sequence. When a grapheme cluster interrupts an unterminated escape (e.g. \x1b[3 followed by ‘), pstate stayed in CSI state, so the next byte was treated as the sequence's final byte and written through the default branch past the truncation point.

A grapheme cluster always returns the parser to the ground state, so this moves the reset to cover all three exit paths.

Before the fix:

Truncate("\x1b[3‘s", 1, "…")          // "\x1b[3…s", width 2
Truncate("7\"(z\x1b[3‘saB'um", 5, "…") // width 6

The fuzz target in the issue failed almost immediately before; with this change StringWidth(Truncate(s, n, tail)) <= n holds across a long fuzz run. Added TestTruncateObeysWidth for the case.

This only touches Truncate (and TruncateWc, which shares truncate()). I left TruncateLeft alone: as noted in the thread it removes n columns from the left, so its result is intentionally total-n wide and the issue's fuzz assertion checks a different invariant for it.