Ensure pre-commit tests by adding actions

[henrykironde]

Add GitHub Actions to enforce pre-commit tests alongside existing app tests

[musaqlain]

@henrykironde I have a proposed solution for replacing the Image.MAX_IMAGE_PIXELS = None line with a tiled raster check, based on your feedback here: #1324 (comment)

Want me to open a follow-up PR that implements it? That way the # to fix marker can be resolved properly.

[vickysharma-prog]

@henrykironde Following up on your tiled raster check suggestion — I've looked deeper into the implementation:

The current global MAX_IMAGE_PIXELS = None (#1326) works but disables PIL's protection entirely. Your original idea of a pre-flight check is cleaner and safer:

import rasterio
import warnings

# PIL default: int(1024 * 1024 * 1024 // 4 // 3) = 89,478,485
PIL_PIXEL_LIMIT = 89_478_485

def _check_large_image(raster_path):
    """Pre-flight check: detect large images and route to optimal strategy."""
    try:
        with rasterio.open(raster_path) as src:
            total_pixels = src.width * src.height
            is_tiled = src.profile.get('tiled', False)
            
            if total_pixels > PIL_PIXEL_LIMIT:
                if is_tiled:
                    warnings.warn(
                        f"Large tiled raster ({total_pixels:,} pixels). "
                        f"Auto-switching to 'window' strategy."
                    )
                    return "window"
                else:
                    raise ValueError(
                        f"Image ({total_pixels:,} pixels) exceeds PIL limit "
                        f"and is not tiled. Convert with:\n"
                        f"  gdal_translate -of GTiff -co TILED=YES input.tif output.tif"
                    )
    except rasterio.errors.RasterioIOError:
        # Non-raster format (PNG/JPG) - continue with default strategy
        pass
    
    return None

Edge cases I've identified:

Non-raster fallback: rasterio.open() fails on PNG/JPG — needs graceful handling (not just crash)
predict_tile dual input: Accepts both raster_path (file) and image (numpy array) — check only applies to file paths
Consistency: download.py (line 22) also sets MAX_IMAGE_PIXELS = None — both should use the same approach
Dual PIL thresholds: Warning at ~89.5M pixels, hard error at ~178M — should we handle these differently?
This would be placed at the start of predict_tile, before any dataset creation — zero memory overhead since rasterio.open() only reads metadata.

I'd like to implement this as a follow-up PR to #1326 with tests. Happy to get started if this direction works!

[henrykironde]

Thanks @musaqlain and @vickysharma-prog for working on this issue. I’m still reviewing how we should handle large files, and once I finalize the approach I’ll update you on the next steps regarding the PIL pixel limit issue and the overall design.