terminal/osc: support iTerm2 OSC 1337 multipart File= transfers

[deblasis]

Follow-up B to #377. The single-shot File= path can only carry images that fit in a single OSC payload (~2 KiB after base64); larger images use the multipart split:

OSC 1337;MultipartFile=options    -- start, options only, no chunk
OSC 1337;FilePart=base64_chunk    -- continue, repeat as needed
OSC 1337;FileEnd                  -- terminator

iTerm2's spec has no session identifier, so transfers are strictly serialized: one in flight at a time.

Wire-up

• New osc.Command.Iterm2MultipartEvent union { start: hints, chunk: bytes, end } emitted one-per-OSC by the parser.
• New Iterm2MultipartAssembler in osc/parsers/iterm2.zig stitches events across OSCs: stashes hints + ArrayList(u8) on start, appends chunks, returns an Iterm2ImageTransmit on end. Owned by the stream handler via a multipart_iterm2 field with deinit on shutdown.
• The MultipartFile arm rejects without inline=1, matching the single-shot File= path; the parser walks options through the same shared parseFileOptions helper so the geometry hints from terminal/osc: honor iTerm2 OSC 1337 File= geometry hints #377 work on multipart too.
• The single-shot iterm2ImageTransmit handler is reused unchanged. After FileEnd the multipart handler frees the assembled buffer immediately after the inner call returns, because synthKittyCommand copies the bytes through base64 decode into a fresh kitty Command buffer.

The payload type on Iterm2ImageTransmit relaxed from [:0]const u8 to []const u8. The single-shot parser still produces a slice into the parser's null-terminated capture buffer, and the multipart assembler now produces a heap-allocated slice without an artificial sentinel. No consumer used the sentinel (simd.base64 takes []const u8).

Error / edge cases

Case	Behavior
FilePart with no active transfer	log.warn, drop chunk
FileEnd with no active transfer	log.warn, no-op
MultipartFile while one is in flight	log.warn, discard previous, start fresh
Accumulated payload above 64 MiB (matches APC kitty cap)	log.warn, drop entire transfer
StreamHandler.deinit	frees the in-flight payload via the assembler

Tests

Parser (6 new):

• MultipartFile with inline+hints
• MultipartFile without inline rejected
• FilePart with chunk
• FilePart with empty value
• FileEnd
• FileEnd= tolerated

Assembler (5 new):

• happy path stitches chunks + preserves hints
• orphan chunk dropped
• orphan end ignored
• overlapping start discards previous
• oversize transfer rejected

Verification

• Windows: zig build test -> 2900/2955 pass, 55 skipped, 0 failures
• Mac (alessandros-macbook-air, arm64, Zig 0.15.2): zig build test-lib-vt -Dtest-filter=iterm2 -> 145/147 pass, 2 skipped
• Linux (ubuntinovm): offline at commit time

Deferred (still tracked from #375)

• JPEG/GIF: needs Ghostty kitty image decoder extension (PNG-only today)
• iTerm2 query/response sequences

Test plan

• zig build test passes on Windows
• zig build test-lib-vt -Dtest-filter=iterm2 passes on Mac
• zig build test-lib-vt passes on Linux (when VM is back online)
• smoke probe-iterm2-multipart.ps1 emits a 3-OSC chunked PNG and Wintty.exe renders it