Embed & Explore folder handling and size-aware auto backend

[NetZissou]

Four independent fixes found while reading the app log for a small image folder, one commit each.

---

a1df5a7 Size-aware auto:

resolve_backend("auto") resolved to cuML whenever a GPU was present, making the size heuristics inside reduce_dim / run_kmeans unreachable from the apps.

Both now use one rule, CUML_AUTO_RANGE, with per-method bounds chosen from measured crossovers of the app's own paths: cuML for PCA from 500 samples and KMeans from 1000, UMAP only from 10000 (subprocess start-up has to be amortized), t-SNE only below 3000 (the collapse-safe exact solver is O(N^2)).

Explicit cuml / sklearn choices are honored unchanged.

Note, the measured numbers behind the bounds are machine-specific , but it should be applicable to general use case.

---

cf71f61 Progress context reports failure

StreamlitProgressContext showed its success message whenever the block exited without raising, so "Embedding complete!" appeared above the "no images" error; the save/repartition sections had the exit message overwrite their error.

New fail(message); all four failure sites use it.

---

87db000 Include subfolders

checkbox in the Embed section (the service already supported recursive listing; the UI never asked for it).

[xxx@xxx Goura_cristata]$ tree -L 1
.
├── eol
└── gbif

---

bdb71e1 No model load for 0 images

generate_embeddings returns before load_model_unified when the listing is empty.

see screenshot above for the verified fix.