OMIM built-in filter: drop the Gene join from the node count query

🤖 Written by Claude.

Follow-up to #1722 — the OMIM built-in filter was the one item #1726 left out ("needs measurement rather than a blind rewrite").

### The filter today

`analysis/models/nodes/node_counts.py` (`get_extra_filters_q`):

```python
elif extra_filters == BuiltInFilters.OMIM:
q = Q(variantannotation__gene__geneannotation__omim_terms__isnull=False)
```

It runs on every node count (aggregated per analysis node), on the `BuiltInFilterNode` queryset, on the node grid when a count label is clicked, and for node column summaries.

### The version claim in #1722 is wrong

#1722 said this matches rows from every `GeneAnnotationVersion` because there's no `version=` filter. It doesn't. Every caller applies the Q to a queryset from `get_variant_queryset_for_annotation_version`, which attaches `AnnotationVersion.sql_partition_transformer` — that rewrites `"annotation_geneannotation"` in the generated SQL to the version's inherits-partition (`annotation_geneannotation_version_2`), the same way it does for `annotation_variantannotation`. Confirmed by dumping the SQL of a node-count aggregate on a dev DB (17M variants):

```
['"annotation_geneannotation_version_2"', '"annotation_variantannotation_version_2"', ...]
```

So there are no cross-version counts to fix, and the fix below does not change any counts.

### What is worth fixing

The join itself. `variantannotation__gene__geneannotation` walks `VariantAnnotation → Gene → GeneAnnotation`, and the `genes_gene` hop carries no filter — it exists only to get from `variantannotation.gene_id` to `geneannotation.gene_id`, which are the same value. `EXPLAIN (ANALYZE, BUFFERS)` on a whole-build count:

```
-> Hash Left Join Hash Cond: (genes_gene.identifier = annotation_geneannotation_version_2.gene_id)
-> Parallel Hash Left Join Hash Cond: (annotation_variantannotation_version_2.gene_id = genes_gene.identifier)
...
-> Parallel Seq Scan on genes_gene (rows=72782, buffers: shared read=5399)
```

A full seq scan + hash build of `genes_gene` (~73k rows, ~42MB of buffers) per count, purely to bridge two columns that already match.

Resolving the gene ids from `GeneAnnotation` as a subquery drops both joins:

```python
gene_annotation_qs = GeneAnnotation.objects.filter(version=gene_annotation_version, omim_terms__isnull=False)
q = Q(variantannotation__gene__in=gene_annotation_qs.values_list("gene_id", flat=True))
```

The subquery hits `annotation_geneannotation_..._gene_id_idx`-backed lookups against a small table (26,390 rows in the version partition, 4,141 with OMIM terms) and is planned once as a hashed SubPlan, not per row.

Filtering `version=` explicitly in Python also means the OMIM filter no longer relies on SQL string rewriting for its version scoping — it stays correct if the Q is ever applied to a queryset without the transformer hook, or on a deployment with `ANNOTATION_GENE_ANNOTATION_VERSION_ENABLED = False` (Shariant).

### Measurements (dev DB, 17M variants, GRCh38, warm cache)

Counting the OMIM filter over a window of variants, two runs each:

| variants scanned | current 3-join | subquery |
| --- | --- | --- |
| 1M | 1.28s / 1.27s | 1.27s / 1.28s |
| 5M | 1.55s / 1.55s | 1.51s / 1.50s |
| 17M (whole build) | 2.49s / 2.46s | 2.19s / 2.19s |

Cold cache the gap is wider (first-ever run of the 1M window: 3.22s vs 1.27s), because the current form has to read `genes_gene` off disk.

This is a modest win, not a #1720-sized one — the honest summary is that it removes work the query never needed, and is most visible when `genes_gene` isn't already cached. Counts are identical: every node of every analysis on the dev DB was counted both ways, 0 mismatches.

### Scope

- `get_extra_filters_q` takes the `AnnotationVersion` instead of the `GenomeBuild` (it needs the `GeneAnnotationVersion`; `genome_build` comes off the annotation version for the classification filters). Callers all have `analysis.annotation_version` in hand.
- Test coverage for the OMIM built-in filter, which currently has none.

Start in analysis/models/nodes/node_counts.py at get_extra_filters_q, then inspect its callers and how each supplies annotation.version. Add coverage for the OMIM built-in filter and verify that counts remain unchanged; the change is done when the Gene join is removed, version scoping is explicit, and the tests pass.

Written by the indexing model from the issue text.