Cut the main-thread cost of opening and re-rendering a visualisation

[NAThompson]

Profiling the h5wasm demo on a 2068x2162 uint16 frame put most of the time between clicking a dataset and seeing pixels in avoidable allocation and shader recompilation. Click-to-figure goes from $211$ ms to $102$ ms.

Per-element allocation: getBounds and getBoundsWithErrors walked values with for..of and .entries(), which allocate per element in V8 and JSC -- scanning a 34 MB Float64Array triggered 4037 young-generation collections where an indexed loop triggers 4 ($3.25\times$ in V8, $1.25\times$ in SpiderMonkey, which already elides most of it). The complex visualisations allocated their outputs with Array.from({ length: n }), whose PACKED_ELEMENTS kind boxes every value written in; a Float64Array and indexed iteration take $4.19$M values from $572$ ms to $160$ ms. Both are tested against naive reference implementations, since hand-written expectations would not pin down code written this way.

Shader recompilation: HeatmapMaterial and GlyphMaterial rebuilt their args object every render, and R3F compares it element by element, so each re-render replaced the ShaderMaterial and relinked its GLSL -- $1.9$ ms in Chrome, $4.7$ ms in Firefox. Both now memoize on what actually appears in the shader source and update the rest in place. useWarmUpWebGL moves the first context's $9$ ms of driver setup into idle time.