ossia: deduce a port's unit from the arity of its value

[jcelerier]

The concepts behind the ossia binding's unit deduction nest: a value with
x, y, z also has x, y, and one with r, g, b, a also has r, g, b. There
was a single xy_parameter overload and a single rgb_parameter one, so every
three- and four-component port was declared to be in position.cart2D, and every
colour in color.rgba.

Each overload now excludes the wider ones, which is how the TouchDesigner binding
already chooses between appendXY / appendXYZ / appendXYZW:

value	unit
.x .y	cartesian_2d_u — unchanged
.x .y .z	cartesian_3d_u
.x .y .z .w	no unit, val_type::VEC4F
.r .g .b	rgb_u
.r .g .b .a	rgba_u — unchanged

xyzw deliberately gets no unit: the only four-component unit in ossia is
orientation.axis, and claiming axis-angle for a generic vec4 control would be
exactly the kind of wrong inference this fixes. The arity is what the conversion
machinery needs there.

Affected in score: TinyObj's Position / Rotation / Scale, Point Tracker 3D's
Anchor, and in the addons SolvePnP's object points. Nothing in tree declares a
three-component colour today, so the rgb_u branch is new ground.

halp_meta(unit, "...")

That declaration went through ossia::parse_dataspace, which knows only the eight
dataspace names — so "midipitch", "distance.mm", "color.rgb" and everything
else a process might actually want to name resolved to an empty unit. It now goes
through parse_pretty_unit, which accepts all of those and answers with a concrete
unit rather than a dataspace with nothing picked inside it. A name that does not
parse leaves the port alone instead of wiping the value type deduced just above.

The only in-tree user is score's Pitch To Value (unit, "midipitch"), whose outlet
consequently gains a real unit (time.midinote) where it previously had none.

The two duplicated blocks in setup_inlets / setup_outlets are folded into one
setup_value_port::setup_port<Field>(), which is also what the test drives.

Testing

Covered by a new test on the score side (Avnd_port_unit_Test, 16 assertions),
since that is where the binding is instantiated: arity for xy / xyz / xyzw and
rgb / rgba, a declared unit resolving, a declared unit winning over the shape of
the value, and a bare dataspace name giving its neutral unit. Reverting the
overload constraints fails 4 of its assertions; reverting the parse_pretty_unit
change fails 3 more.

A full score build is clean — every avnd process in the tree recompiles against
the new overloads.

Ordering

Part of a three-repo change. This one stands alone; ossia/libossia and ossia/score
carry the rest, and the score PR needs all three.

🤖 Generated with Claude Code

https://claude.ai/code/session_0197hSihNtB1bPw9WDUV3SR2