fix(dynamic-atlas): use dynamic batch size to prevent glyph clipping at large font sizes

[kofany]

Summary

Fixes #67

The OffscreenCanvas used for glyph rasterization has a fixed height of 1024px. When using larger font sizes (e.g., 30px resulting in ~60px cell height), the previous hardcoded batch size of 32 glyphs would exceed the canvas height (32 × 60px = 1920px > 1024px), causing glyphs beyond the canvas boundary to be clipped/empty.

This was the root cause of characters like ~, -, <, > disappearing at larger font sizes - they appeared late in the ASCII batch and were rendered beyond the canvas boundary.

Why ~ (tilde) disappeared

ASCII ~ is code 0x7E (126). In the ASCII glyph upload, it's at position 94 (126 - 32). With batch size 32:

• Batch 0: glyphs 0-31 (space to ?)
• Batch 1: glyphs 32-63 (@ to _)
• Batch 2: glyphs 64-95 (` to ~)

Tilde at local position 30 in batch 2 means Y = 30 × 60px = 1800px - well beyond the 1024px canvas!

The fix

Adds CanvasRasterizer::max_batch_size() which dynamically calculates the maximum batch size based on cell height:

pub(crate) fn max_batch_size(&self) -> usize {
    (OFFSCREEN_CANVAS_HEIGHT / self.cell_metrics.padded_height).max(1) as usize
}

This keeps the batch size knowledge encapsulated in the rasterizer where it belongs, and DynamicFontAtlas simply calls self.rasterizer.max_batch_size() instead of using a hardcoded value.

Test plan

• Tested with font sizes 25px, 30px, 35px - all ASCII characters render correctly
• Verified emoji and CJK characters still work (they use upload_pending_glyphs which also uses dynamic batch size now)

[junkdog]

i like the approach but i think it's better to rescale the offscreen canvas to fit 32 glyphs, to avoid creating too many (tiny) batches.

[kofany]

i like the approach but i think it's better to rescale the offscreen canvas to fit 32 glyphs, to avoid creating too many (tiny) batches.

I will try this approach maybe during weekend. Thanks for pointing direction :)

[kofany]

@junkdog Check it now please :)

[junkdog]

looking good, thanks!