Working toward composing things like flocking from attribute math alone, without custom shaders, and two things were missing.

• Vector length ops. MAP now has NORMALIZE (to length=) and LIMIT (min_length=/max_length=). Both work on the whole vector; zero stays zero.
• Broadcasting. Like numpy's (N, 1) * (N, 3): a 1-component b in COMBINE now applies across every component, so you can scale by a per-particle value or multiply by a 0/1 mask.

p.apply(MAP, "velocity", op=LIMIT, max_length=10)
p.apply(COMBINE, "velocity", "speed", op=MUL)

Positional operands come in the next PR.