Use exact solver for cuML t-SNE to avoid Barnes-Hut collapse

[NetZissou]

cuML's default Barnes-Hut t-SNE collapses to a ~1D diagonal line on near-homogeneous embeddings, we're able to reproduce this on the Darwin's finches BioCLIP 2 embeddings. The "exact" solver fix this easily.

Force method="exact" for the cuML t-SNE path. It's O(N^2) but fine at the interactive scale this app runs; a faster Barnes-Hut-with-degeneracy-guard can be revisited later.

Closes #40

[NetZissou]

@egrace479 this fix won't affect non-cuda accelerated sessions.

I've validated the changes on pitzer GPU.

[2026-06-18 08:56:36] INFO [shared.utils.backend.check_cuda_available:52] CUDA        
  available via PyTorch: Tesla V100-PCIE-16GB
  [2026-06-18 08:56:42] INFO [shared.utils.backend.resolve_backend:105] Auto-resolved   
  reduction backend to cuML (GPU: Tesla V100-PCIE-16GB)                                 
  [2026-06-18 08:56:42] INFO                                                            
  [apps.precalculated.components.sidebar._run_projection:620]
  ============================================================
  [2026-06-18 08:56:42] INFO
  [apps.precalculated.components.sidebar._run_projection:621] PROJECTION START
  [2026-06-18 08:56:42] INFO
  [apps.precalculated.components.sidebar._run_projection:622] Device: Tesla
  V100-PCIE-16GB (CUDA: Yes)
  [2026-06-18 08:56:42] INFO
  [apps.precalculated.components.sidebar._run_projection:623] Backend: cuml (requested:
  auto)
  [2026-06-18 08:56:42] INFO
  [apps.precalculated.components.sidebar.extract_embeddings_safe:517] Extracting
  embeddings from DataFrame: 677 rows
  [2026-06-18 08:56:42] INFO
  [apps.precalculated.components.sidebar.extract_embeddings_safe:526] Extracted
  embeddings: shape=(677, 768), dtype=float32
  [2026-06-18 08:56:42] INFO
  [apps.precalculated.components.sidebar._run_projection:631] Records: 677 | Dim: 768 |
  Extracted in 0.00s
  [2026-06-18 08:56:42] INFO [shared.services.clustering_service.run_dim_reduction:104]
  Dim reduction: samples=677, features=768, method=TSNE, backend=cuml, seed=614
  [2026-06-18 08:56:42] INFO [shared.utils.clustering.reduce_dim:165] Dimensionality
  reduction: method=TSNE, samples=677, features=768, backend=cuml
  [2026-06-18 08:56:42] INFO [shared.utils.clustering._prepare_embeddings:141]
  [reduce_dim] Prepared embeddings: 677 samples, 768 features, dtype=float32,           
  L2-normalized (input norms: min=40.57, max=54.16, mean=44.74)
  [2026-06-18 08:56:42] INFO [shared.utils.clustering.reduce_dim:181] Using cuML
  backend for TSNE
  [2026-06-18 08:56:43] INFO [shared.utils.clustering.reduce_dim:188] Dimensionality
  reduction completed in 0.72s
  [2026-06-18 08:56:43] INFO [shared.services.clustering_service.run_dim_reduction:112]
  Dim reduction complete: (677, 2) in 0.75s

[egrace479]

LGTM. We can revisit and explore other speed-up options if needed.