feat(host): opt-in ADVANCED_OK buffer-aware look-ahead streaming (deferred)

[CameronBrooks11]

Deferred — evidence-gated

Streaming is currently plain send-one-await-ok pacing (host.py:232-250) — the
lowest-common-denominator handshake every Marlin supports. ADVANCED_OK adds
ok N<line> P<planner_free> B<blocks_free> (gcode/queue.cpp:251-263), enabling
buffer-count look-ahead. Prior art (design doc Lane C): the ecosystem treats this as
a fragile opt-in — only OctoPrint's third-party BufferBuddy plugin uses it; core
OctoPrint, Cura, and printrun all deliberately ignore P/B.

Proposed approach (only when a real workload needs it)

Behind an explicit flag, and only when profile.advanced_ok is confirmed
(firmware proven to emit P/B), allow sending ahead until fewer than B commands
are in flight (BufferBuddy-style) to avoid planner underrun on small segments. Stay
one-outstanding by default. Never auto-enable.

Acceptance criteria (when/if taken)

• Look-ahead is off by default and gated on confirmed ADVANCED_OK fields.
• One-outstanding behavior unchanged when the flag is off or fields absent.

Evidence

_reference/Marlin/Marlin/src/gcode/queue.cpp:251-263, marlin_host/protocol.py:89-95,
marlin_host/host.py:232-250; OctoPrint BufferBuddy / OctoPrint/OctoPrint#2834.
Design: working/dialect-and-mock-investigation.md (Later).

Part of #5.