feat: data-oriented input sources

[kvvasuu]

Reworks input into small per-frame data sources, so cameras can be driven by the mouse, touch, wheel and pinch, or by any data.

Changes

• InputAxis.setValue jumps to a value without easing.
• VirtualCamera.inputAxes collects the axes of all its pieces. <InputController> placed directly in <VirtualCamera> uses them.
• Mappings are optional per source, and the new single-axis mappings are wheel and pinch.
• Mouse, touch, wheel and Safari gesture input now come from internal sources in src/input-sources/. Lint keeps that folder free of klipp imports. InputSystem is a thin layer over them.
• New docs page "Input from data" and an "Input" examples category.

Fixes

• frameloop="demand" wakes on every input event, including the wheel.
• Line- and page-based wheel deltas are converted to pixels.
• The canvas gets its inline touch-action / user-select back after disconnecting.
• Held buttons and fingers are released on blur and lost pointer capture.
• iOS touch pinches are no longer counted twice.
• The page keeps its scrolling and pinch zoom unless wheel / trackpadPinch is mapped.

Breaking

• InputAxisController → InputControllerDom, lastInput → input.
• Config types (InputControllerConfig, InputSourceMapping, …) are exported from @kvvasuu/klipp, no longer from /dom.
• ConsumedInput.gestureZoomDelta is now the change in ln(scale).
• ConsumedInput.touchPinchDelta, gestureZoomDelta and wheelZoomDelta are replaced by one pinchDelta, the change in ln(scale) from fingers or a trackpad.

Benches

Input benches are rewritten around a fake DOM. All benches warm up before yield, since labs otherwise decides on batching from cold first calls and reports inflated times.