Visualizer: cache idle plates and send method signatures once per class

[BioCam]

Motivation

On a busy deck the visualizer became expensive in two places, both of which are felt most on low-power hosts such as a Raspberry Pi: a live update stutters during a run, and the first connect is slow. This PR addresses both, with no change to what the user sees.

Idle-plate caching (runtime cost)

A whole-layer Konva redraw costs in proportion to the number of shapes on the layer, so after the recent change that forwards volume updates to the visualizer, a single well update repaints every other plate and tip rack on the deck. On a six-plate deck, updating one plate's 96 wells measured about 4.6 ms per frame.

While a plate or tip rack is idle its group is cached to an offscreen bitmap, so it blits in one operation instead of re-rendering all its wells or tips. When a state update arrives for a plate its cache is cleared so the change renders live, and it is re-cached once it has been idle briefly. The group hierarchy is untouched, so transforms, rotation, moves, and hit detection are unaffected. The same six-plate update drops to about 0.55 ms and, more importantly, stays flat as the deck grows.

The cache resolution tracks the on-screen device pixels so an idle plate stays as sharp as a live one, and is rebuilt sharper when the user zooms in. Sidebar hover highlights and per-item visibility toggles clear the affected plate's cache so their changes show immediately, then re-cache when idle.

Method signatures once per class (startup cost)

The initial resource message attached the full list of Python method signatures to every resource, so a plate repeated the same list on all 96 of its wells. This is the largest single cost at connect, and the browser must parse the whole blob before the first paint.

Signatures are identical for every instance of a class, so they are now sent once in a registry keyed by resource type; the browser attaches the matching signatures by type when it builds the tree. On a six-plate, six-tip-rack deck this shrinks the payload from about 2.7 MB to 0.77 MB (roughly 3.5x smaller). The UML method panel is populated exactly as before.

Testing

Existing visualizer tests pass (updated for the new message shape). Verified through a headless browser driving the real websocket: the cache lifecycle is correct (idle plates cached, active plate live, re-cached after idle, no stale bitmap); hit detection still works over cached plates; zoom, hover-highlight, and per-item visibility behave correctly; a real liquid-handling flow (tip pickup, aspirate, dispense) renders correctly with both changes active; and method signatures attach in the browser via the registry.