perf(webgl2): reduce USM shader from 9 to 4 texture fetches (-30% draw GPU)

[Endymi0n74]

What

The unsharp-masking path of the WebGL2 fragment shader (clarity-boost.fs) sampled all 9 texels of the 3×3 neighborhood to compute the gaussian blur. This PR replaces those 9 fetches with 4 bilinear samples at the edge midpoints (±0.5 texel) — the sum of the 4 mid-edge samples is exactly the 3×3 gaussian [1,2,1;2,4,2;1,2,1]/16, so the corner samples a,c,g,i are no longer needed in the USM branch. The CAS path is untouched.

Why

USM is the default processing filter and runs once per pixel on the render hot path. At 640×360 the draw call is ~10 µs; the gain is strictly GPU-side (same GL counters: 1 texSubImage2D, 1 drawArrays, 0 uniform/frame) and scales with pixel count.

Measurement (seed 42, 640×360, WebGL2)

Variant	Draw GPU	Δ
before (9 fetches)	10,24 µs	—
after (4 taps)	7,17 µs	−30 %

Identical result on the 3 measured passes. At 1080p (9× the pixels) this extrapolates to ~30 µs/frame.

Visual equivalence

Pixel-diff over fine text (6 cases: USM default / strong / off × quality on / off):

• sharpness off: bit-identical output
• sharpness on: max channel diff ±1 ULP on ≤ 0,002 % of pixels (bilinear rounding at the midpoint samples)

Caveat — WebGPU

The same 4-tap idea was prototyped on the WebGPU path and measured slower on Dawn/D3D12 (4 bilinear taps = 16 underlying texel fetches vs 9 direct fetches, +3-7 %). The gain is specific to the WebGL2/ANGLE stack, so this PR only touches clarity-boost.fs.