fix(rio-vt): close synchronized updates whose markers arrive split

[raphamorim]

Fixes the synchronized-update (mode 2026) bug reported in #1894: a BSU partway through a chunk armed the timeout but let the rest of the chunk bypass the sync buffer, so an ESU split across two reads was never recognized and a one-frame TUI froze on the previous frame until the timeout.

Approach

Instead of #1894's byte-pattern scan over every chunk (which measured 18-26% slower on escape-dense input), this ports the design upstream vte 0.15 chose for the same bug in alacritty: the parser itself stops right after dispatching a BSU. Perform gains a terminated() hook, and the new Parser::advance_until_terminated returns how many bytes it consumed; the Processor routes the remainder to the sync buffer. The cut is exact for every BSU form (single-param, multi-param like \e[?2026;1004h, the DCS =1s form, and any of them split across reads), because it keys off the dispatch rather than a byte pattern. Termination can only flip on a CSI dispatch (parser lands in Ground) or a DCS hook (parser lands in DcsPassthrough), so the check sits at the entry of those two arms only: no per-byte or per-iteration cost.

For reference, ghostty and kitty cannot have this bug class: both parse every byte through the state machine and implement 2026 as render gating (ghostty skips renderer frames with a 1s watchdog; kitty's screen_pause_rendering snapshots the screen, 2s expiry). That architecture needs the renderer's cooperation, which rio-vt cannot assume: it is a library whose embedders (librio C/Swift hosts, wasm) get sync correctness from the byte-buffer design for free. Within that design, the parser-driven cut is the fix upstream alacritty shipped.

Performance

Performer takes a const REPLAY parameter distinguishing the sync-replay instantiation from the PTY hot path. This matters: a naive second advance loop gives every run helper (advance_ground, advance_csi_param_run, ...) two callers, LLVM stops inlining them, and escape-dense throughput drops 5-10% (found via samply profiles: advance_ground went from fused into the loop to a standalone symbol). With the replay path as its own monomorphization, every helper is single-caller again and inlining is restored by construction.

Benchmarked with the vt_input suite, both binaries prebuilt and run back to back against a fresh baseline (stale baselines drift several percent thermally over an hour, verified with a same-code control):

corpus	change vs main
plain	+0.13% (within noise)
ansi_mixed	+0.36% (p = 0.11, not significant)
ansi_bare_csi	+0.06% (p = 0.89)
truecolor_unique	-0.37% (within noise)

Tests

533 rio-vt tests pass. Five new ones: mid-chunk BSU with split ESU (CSI and DCS forms), multi-param BSU (which a byte-pattern scan cannot cut on), split BSU with a tail in the second chunk, and an exhaustive chunking-equivalence test that replays both streams at every chunk size from 1 upward and asserts identical results.