Streaming mask support: avoid f32 path and full materialization for pixel-wise masks

Summary

Masks in zenresize (via with_mask()) currently force the f32 pipeline path. Since all mask types (RoundedRect, LinearGradient, RadialGradient) compute values per-pixel per-row, there's no fundamental reason they can't work in the i16 paths too.

Additionally, zenfilters' Masked filter wrapper operates on full OklabPlanes buffers, but for row-local inner filters (exposure, contrast, saturation, tone curves — anything that doesn't need spatial neighborhood), the save→apply→mask→blend cycle could stream row-by-row:

1. Save original row
2. Apply filter to row (row-local filters only)
3. Compute mask values for that row (all mask types are y-addressable)
4. Blend: result = original * (1 - mask) + filtered * mask
5. Emit row

What this enables

• i16 mask path in zenresize: per-pixel lerp in i16 space ((a * (256 - m) + b * m + 128) >> 8 for u8 masks). Avoids f32 conversion overhead for the common case of rounded corners + solid background → JPEG.
• Streaming masked filters in zenfilters/zenpipe: color grading + gradient mask could stream without materializing the full image. Only non-row-local inner filters (blur, warp, sharpen with large radius) would still need full materialization.
• Composable with zenresize's existing streaming pad/crop/orient: mask application slots in after resize, before orient — same position it's in now, just without forcing f32.

Non-goals

• Masks that depend on image content (future content-aware masks) are inherently non-streaming.
• Non-row-local inner filters (blur, warp) still need full materialization regardless of mask streaming.

Relevant code

• zenresize/src/streaming.rs — with_mask() forces f32 path
• zenblend/src/mask.rs — MaskSource trait, MaskSpans::for_row(y, width)
• zenfilters/src/masked.rs — Masked wrapper, full-image save→apply→blend
• zenfilters/src/filter.rs — Filter::is_neighborhood() / neighborhood_radius() — already distinguishes row-local from spatial filters

Start with zenresize/src/streaming.rs to trace how with_mask() selects the f32 path, then read zenblend/src/mask.rs and MaskSpans::for_row(y, width). Review zenfilters/src/masked.rs and zenfilters/src/filter.rs to distinguish row-local from neighborhood filters. Done means i16 mask processing and row-by-row masked filtering work without full materialization where the stated constraints allow it.