Refactor LineageTree to SciPy CSR sparse array representation and array-level simulation (#1000)

[aarmey]

Resolves #1000

Overview

This PR refactors the core tree representation and algorithm recursions in tHMM:

1. SciPy CSR Sparse Array Lineage Representation:

   • Replaced node-based pointer trees with a scipy.sparse.csr_array adjacency representation (lineage.tree) storing directed parent $\to$ daughter edges.
   • Vectorized parent-daughter edge indexing using CSR indptr and indices.
   • Vectorized transition likelihood scoring (log_T_score) in tHMM.py using flat edge tensors, yielding a ~60x speedup.
   • Vectorized M-step transition probability estimation in lineage/HMM/M_step.py via np.einsum.

2. Array-Based Tree Simulation & Censoring:

   • Replaced CellVar heap object instantiations in LineageTree.rand_init with pure array generation (tree, obs, states).
   • Implemented vectorized temporal and fate censoring (censor_lineage_gamma and censor_lineage_gaphs) operating directly on CSR arrays and observation matrices.
   • Purged dead legacy CellVar helper functions (apply_censoring, basic_censor, assign_times, fate_censor, time_censor).

3. Documentation & Validation:

   • Updated docs/tHMM-getting-started.rst to document the array and CSR sparse data model.
   • All 43 test suites passing cleanly with full typing (ty check) and linting (ruff check) validation.