fix(render): the motion-vector pass waits for a reader

[lobinuxsoft]

~2 ms per frame, measured, with no quality cost. The first item of the graphics Phase 0 to produce code.

Measured on the OneXFly, many_lights.scene, temporal resolve off:

motion vectors   1.994 ms   ← of a 20.5 ms GPU frame, budget 13.9

taa.wgsl is the only shader that binds the buffer, so with the resolve off the pass writes a full-resolution target nobody samples. The comment above it explained why it runs in every debug mode — true, and silent about whether anything reads the result.

• The predicate is needs_motion(), not self.taa_enabled at the call site: FSR (feat(render): FSR over the FidelityFX C ABI (FFI), behind the upscaler trait #536) becomes the second consumer of the same condition, so it extends one gate instead of adding a second.
• The history the vectors are computed from is maintained by the scene upload, not by this pass — so turning the resolve back on mid-session gets correct vectors on its first frame. Documented on the predicate.

On the test, because the obvious one is unsound. "The motion vectors scope is absent" cannot be asserted in a shared test binary: puffin's scope_delta is a delta, the first test to register a name drains it, and gpu_scopes.rs already documents that asserting absence there is asserting which test ran first. So this reads the texture — a fresh wgpu texture is zeroed — in its own binary, and the resolve-on half is what stops the resolve-off half from passing vacuously (a gate that skipped the pass unconditionally would pass one and fail the other).

🔴 Verified failing with the gate reverted: 147684 of 160000 bytes are non-zero, so the pass is still running for nobody.

temporal_motion, half_rate_shading and compute_shading_parity all still pass. Refs #481.