feat: make tiling the default input_prep, with ViT patch-grid safety guards

[Dinghye]

Summary

Changes the default input_prep policy from "resize" to tiling, so that large on-the-fly inputs preserve native resolution by default instead of being silently downsampled to the model's image size.

• New default. input_prep=None is now the "unset / use package default" sentinel and resolves to "tile". This default is applied consistently across get_embedding, get_embeddings_batch, Model(...), and ExportConfig. "resize" and "auto" remain available as explicit opt-ins.
• Model-aware safety defaults. For image-level ViT patch-token-grid models (remoteclip, scalemae, satmae, satmaepp, satmaepp_s2_10b), an unset/"auto" input_prep resolves to "resize" instead of tile, because their tiled patch-token grids are not a seamless spatial mosaic and can show stitching seams. Requesting grid output on these models emits a warning; explicit "tile" still works but warns and records seam-risk metadata. The resolved embedding records the model policy in meta["input_prep"].
• Soft + hard tile budgets. max_tiles default raised 9 → 64 and is now a soft threshold (requests above it still run but warn that they may be slow / memory-heavy). A new max_tiles_hard ceiling (default 1024, clamped >= max_tiles) raises to guard against runaway-area mistakes.
• Graceful degradation. Because tiling is now the default, the tile path no longer hard-fails when a prerequisite is missing (no provider backend, no sensor, embedder doesn't accept prefetched CHW, or no derivable tile size). It degrades to a plain resize call and records resolved_mode="resize" in meta, so basic usage never breaks and the fallback stays auditable.
• Reproducibility metadata. Every embedding now carries a versioned meta["input_prep"] block (INPUT_PREP_VERSION = 1) recording requested vs. resolved mode (and tiling layout / counts when tiled), regardless of which code path produced it.
• remoteclip pooled/grid consistency fix. The single path now uses encode_image (canonical 512-d CLIP embedding) so pooled is identical across single, batch, and tiled paths instead of being a 768-d raw-token mean on the single path. grid extraction was rewritten to read open_clip's forward_intermediates patch grid. See docs/models/remoteclip.md.

Why: the previous "resize" default quietly discarded spatial resolution for larger ROIs, and tile mode was easy to trip into hard errors. This makes the high-fidelity behavior the default while guarding the cases (ViT patch grids, huge areas) where naive tiling misleads.

Testing

How did you verify the change?

• I ran the relevant tests locally.
• I updated docs for user-facing behavior changes.
• I updated CHANGELOG.md for user-facing API, model, semantic, or installation changes.
• This PR does not need a changelog entry.

New/updated coverage: tests/test_model_aware_input_prep.py (default resolves to resize with metadata for ViT-grid models; explicit tile allowed but marked experimental), plus updated test_input_prep_tiling.py, test_specs.py, test_api.py, test_batch_overrides_all_models.py, test_combined_flow.py, test_export_batch.py, and test_scalemae_preprocess_alignment.py.

Notes

• Behavior change (not strictly backward compatible). Code relying on the old "resize" default will now tile by default; results for large ROIs change shape/values. Pin input_prep="resize" to keep prior behavior. Documented in CHANGELOG.md under Changed.
• The ViT patch-grid models intentionally opt out of the new default and stay on resize — pooled output is the recommended path for those.
• Docs updated: choosing_settings.md, concepts.md, api_embedding.md, api_specs.md, and the per-model pages for the affected ViT models.