fix(art): detect actual image format and save with correct extension

[VintLin]

Problem

generate may save image bytes with a mismatched extension (e.g., JPEG bytes written to a .png path). This can later cause MIME-type validation failures in downstream readers/tools that trust the extension.

Additionally, reference-image MIME type selection in nano-banana-pro was extension-based only, which can also break if file suffix and bytes diverge.

Changes

• Added magic-byte image format detection (png, jpeg, webp, gif).
• Added smart output writer that:
  • detects actual output format
  • rewrites filename extension when needed
  • returns the final output path
• Updated all generation backends (flux, nano-banana, gpt-image-1, nano-banana-pro) to use smart writer.
• Updated nano-banana-pro reference-image MIME resolution to:
  • prefer content-based detection
  • fallback to extension when detection is unavailable
• Tracked the actual generated output path through post-processing (remove-bg, add-bg, thumbnail) so follow-up steps apply to the real file location.

Validation

• Type/path-level validation by tracing all generation paths and ensuring post-processing uses generatedOutput returned by generator functions.
• Verified no behavior change when format already matches requested extension.
• Verified mismatch path now saves with correct extension and emits a warning.

Fixes #633

[kaimagnus]

Thank you for this excellent fix, @VintLin! You correctly identified the core issue — the API sometimes returns a different image format than requested.

We've addressed this in commit 0fc1931 with a similar approach (magic byte detection, automatic extension correction, and proper MIME type handling throughout the pipeline). Closing in favor of that fix. Your analysis of the problem was spot-on and really helped confirm the right approach!