Analysis Export with Provenance

## Motivation

Neither app can export the analysis table today: `embed_explore` only saves/repartitions images, `precalculated` has no download at all. Compute results (2D projections, KMeans runs, and soon annotations from #47) live only in the user's Streamlit session.

Exporting **projection coordinates as a snapshot** is also the only reliable way to reproduce a plot: cuML t-SNE (default solver) is not deterministic even with a seed (see #46 / README reproducibility section), so a "re-run with seed" recipe cannot reproduce a GPU t-SNE view but a coordinate table always can.

## Scope

One export action, shared by both apps (`shared/components`), producing two files:

### 1. Table (CSV) flat, one row per record
| group | columns | notes |
|---|---|---|
| key | `uuid` (precalculated) / `file_path` (embed_explore) | join key back to source data |
| projections | `_x`, `_y` per projection run this session | tag = method + seed, e.g. `tsne_s614`, `umap_s614`, `pca` |
| clustering | every `KMeans (k=…)` column | as in `df_plot` today |
| annotations | every annotation column (#47) | `"N/A"` = unlabeled |
| metadata | original metadata columns | **toggle "include original metadata"**, default on |

### 2. Provenance sidecar (JSON)
- app, emb-explorer version, library versions (sklearn / umap-learn / cuml), export timestamp
- source dataset (parquet name or image folder) + record count + data hash
- per projection tag: method, backend, seed, key params (perplexity / n_neighbors), run timestamp
- per KMeans column: k, backend, seed
- per annotation column #47: projection tag + KMeans column on screen when created / last edited, timestamps