Switch default optimizer to CG with multi-start

[Marius1311]

Summary

Switch the default GPCCA rotation optimizer from Nelder-Mead to conjugate gradient (CG) with 10 random restarts. This makes compute_macrostates() scale to large numbers of macrostates (m > 15) where Nelder-Mead becomes impractical.

Changes

• Default optimizer: "Nelder-Mead" → "CG"
• Default n_starts: 1 → 10 (with perturbation_scale=0.1)
• New parameters on compute_macrostates(): n_starts, perturbation_scale, seed — these were already supported by pyGPCCA but not exposed in CellRank
• fit() method: Now forwards optimizer parameters to compute_macrostates() (previously always used hardcoded defaults)
• .gitignore: Added uv.lock and .github/prompts/

Benchmark results

Validated on pancreas (2,531 cells) and bone marrow (5,780 cells) datasets:

m	NM 1× crispness (time)	CG 10×0.1 crispness (time)
5	0.513 (0.1s)	0.477 (0.2s)
10	0.234 (1.6s)	0.243 (0.4s)
15	0.123 (13.7s)	0.221 (2.2s)
50	infeasible	0.063 (30.0s)

• CG 10×0.1 matches or beats NM on crispness at m ≥ 8
• High membership agreement: mean Pearson r ≥ 0.94, Jaccard ≥ 0.70
• All CG solutions are non-degenerate (no empty clusters)
• Extrapolated: CG 10×0.1 at m=50, n=1M ≈ 3.3 hrs (NM infeasible)

Dependencies

Requires pyGPCCA from msmdev/pyGPCCA branch feature/gradient-optimization which implements the analytical Jacobian and multi-start SO(k) rotation perturbation.