feat: metadata filtering (#45) and force-categorical toggle (#35)

[RaredonLab]

Closes #35
Closes #45

Both issues are the same question asked twice — "what kind of thing is this metadata column?" — so they share one module, readers/metadata_filter.py.

#35 — treat as categorical
Integer cluster IDs from Seurat were routed to a viridis gradient by dtype.
color_values() and edge_color_values() now take an explicit categorical
override, and category labels sort numerically so cluster 10 follows 2.

_color_values_meta moves to the base class. All six readers carried a
near-identical copy and the copies had drifted: CosMx filled NaN with ""/0
where the others dropped it, and only some sorted with key=str. A reader now
supplies only _metadata_frame().

The frontend stops guessing the type from the schema dtype — it could not,
since the rule also depends on cardinality. The old guess disagreed for
exactly the columns this issue is about: the canvas drew discrete colours
while the panel showed a gradient with two sliders that did nothing. Panel 0
now reports the type the backend returned. That also removed the duplicate
color-values fetch both legends were making for themselves.

#45 — select cells and edges by metadata
A MetadataFilter is a categorical allowlist or an inclusive numeric range,
resolved server-side and applied BEFORE sampling. That ordering is the point:
both queries sample on the server, so a client-side filter would leave a
fraction of a subset — a cluster holding 5% of cells at a 10% sample would
draw 0.5% of the tissue.

cell_boundaries() takes cell_ids in all five readers. query_grouped() takes
cell_ids and edge_filter; an edge survives only when both endpoints do, since
a half-outside edge runs off to a cell that is not drawn. Large id sets go
through a registered DuckDB relation rather than an IN list.

An unknown column raises 400 rather than rendering everything under an
apparently-active filter.

Also fixed, because the filter exposed it: useCellBoundaries picks its auto fraction from the previous fetch's total, which a filter invalidates, and nothing else triggered a refetch — so the layer sat showing a tenth of an already-small subset until the user panned. It now recalibrates once.

Guard extended to 191 probes across 7 datasets, covering both features on every platform. The demo fixture gains a 12-level seurat_clusters column so the repo's own data reproduces #35.