vello_cpu: Fix Color and Luminosity blends turning some grays black

[zlstringham]

Fixes #1999.

clip_color divides by l - n when n < 0 and by x - l when x > 1, without checking the denominator. After set_lum, a gray has three equal components, so l == n == x. When rounding leaves them just below 0 (or just above 1), the division is 0/0. The NaN survives apply_alpha, even over a transparent backdrop, and packs to black. With Mix::Color and Mix::Luminosity, some gray levels come out black in both pipelines.

This PR clips only where the denominator is positive: n < 0 && l - n > 0 and x > 1 && x - l > 0. For a neutral color, the clipped value's limit is the color itself, so skipping the clip loses nothing. Skia (l != mn, l != mx) and pixman (FLOAT_IS_ZERO(l - n)) guard the same case.

vello_gpu_shaders/shaders/helpers/blend.wesl and research/vello_shaders/shader/shared/blend.wgsl have the same unguarded clip_color and get the same guard. I couldn't make the GPU path produce black pixels on Metal, but WGSL gives an indeterminate value for a division without a finite result, so the guard still helps there.

Testing:

• New unit test non_separable_mix_of_neutral_colors_is_finite checks that mix stays finite for Hue/Saturation/Color/Luminosity with every gray at 50% alpha over a transparent backdrop and over every opaque gray. It fails without the guard and passes with it.
• cargo test -p vello_cpu passes.
• cargo test -p vello_tests --release passes on CPU and on GPU (Metal, Apple M3 Max), with no snapshot changes.
• cargo test -p vello_gpu_shaders --all-features passes. vello_gpu_shaders and vello_shaders build, and their build scripts validate the WGSL with naga.
• The reproducer from the issue now reports 0 black pixels for every mode in both render modes.
• cargo fmt --check, typos, the CI clippy commands for vello_cpu (including no_std and wasm32), and the 1.89 MSRV check are clean.