Make MackieBackend an adapter over the shared C translator

[alphonsom]

The Mackie Control translation existed twice: once in MackieBackend for the
host bridge, once in src/mcu/c8_mcu.c for the dongle firmware. Two copies of a
protocol mapping drift, and they drift silently — each stays correct against
its own tests while producing different bytes on the wire.

MackieBackend now owns no translation logic. It holds a c8_mcu_t, feeds
surface input into it, and routes its two output callbacks to the MCU port and
to the surface.

Deleted from the C++ side: the MCU note map, the Bank/Nudge/Zoom nav group, the
LCD SysEx parser, the V-pot ring and meter encoding, and the selected-track set.
213 lines removed against 290 added, and most of the addition is the test.

The C module survives because it is the tested one, and because the firmware
compiles that exact file rather than a copy.

Two consequences worth being explicit about

Feedback gains raw(), a passthrough for back-ends that emit device bytes
themselves. The translator already produces Command|8 wire format, so routing
its output through the named encoders would mean decoding it only to encode it
again.

The adapter costs one re-encode. Surface decodes device bytes into Events,
Controller hands those to the on_* methods, and they rebuild the bytes the
device sent. That is a genuine inelegance. It is kept because the alternative —
a byte-level path from Surface straight into the translator — touches every
Surface backend, and this is the code path that only just started working on
Windows. A follow-up, not a bundled change.

The re-encode is exhaustively verified

tests/test_mackie_roundtrip.cpp proves it rather than sampling it:

• all 8192 fader CC/value combinations (8 strips x 8 low-bit groups x 128 values)
• both encoder detents on all 8 strips
• every note/velocity pair that decodes to a button (>15000)

each driven through decode_* and back, with the bytes required to match
exactly. The domains are small enough that guessing would be a choice.

A lossy round trip would shift control values by a least-significant bit or
two — a fader that never quite reaches unity, an encoder that stops responding —
which is precisely the kind of fault that gets blamed on hardware.

Verified

All 5 test suites pass. command8-mackie builds and fails cleanly with the
right message when no MCU port exists. The dongle firmware still builds against
the unchanged c8_mcu.c, and its own host tests pass.