fix(motion): let counters land, float_3d honour its timing, and pans dissolve

[LeadcodeDev]

Three motion defects that all read the same way on screen — movement that ignores what the scenario asked for.

Counters never landed. The ramp was hardcoded to stretch from start_at to the end of the scene, so the figure reached to on the very last frame and the viewer never got to read the number they had just watched count up. New optional counter.duration (seconds, measured from start_at) makes the count finish early and hold. Unset, the old behaviour is preserved exactly. A duration outlasting the scene is honoured rather than clamped — the count simply never arrives, which is what was asked for; a zero or negative one falls back to the scene. Five unit tests pin the ramp, including the two edge cases.

float_3d ignored both delay and duration. Its keyframes were pinned to literal 0.0 / 0.5 / 1.0 seconds, so every floating element in a scene ran one shared 1-second cycle in lockstep no matter what the scenario declared. A row of cards bobbing in unison reads as a dance; the same cards on different phases read as depth, which is the entire point of the preset. The cycle now spans delay..delay+duration, so duration sets the period and delay shifts the phase.

Camera pans slid the foreground as a solid slab. The outgoing plane now dissolves and the incoming one materialises across the pan. Both curves are pinned at their own end — fg_a fully opaque at t=0, fg_b at t=1 — because a transition frame sits directly against a normal frame at each junction and any alpha short of 1 there is a visible step. Mirrored exponents rather than a plain crossfade keep both planes at 67% through the middle, so the frame never washes out half way. Two tests cover it: the junction no-op invariant, and the mid-pan floor.

Known gap

PresetConfig.amplitude is added and read by the preset expansion, but nothing can currently set it. AnimationTiming — the struct scenarios actually deserialize into — has no amplitude field, and to_preset_config() hardcodes amplitude: None, which is the only construction site outside Default. So config.amplitude.unwrap_or(12.0) always yields 12.0 and the knob is dead from a scenario author's point of view. delay and duration do land, since those come from AnimationTiming. Wiring it through is a field on AnimationTiming plus one line in to_preset_config; left out of this PR rather than smuggled in.

Verification

cargo test --workspace green. cargo check --workspace clean.