Fix/bicv alignment and diagnostics

[fishidaho]

The goal of this PR is to fix #546 and close #547 as well as #548.

---

#546 — _bicv_trial misaligned Z against the rows it is regressed on

The held-out gene regression built its design matrix by concatenating per-condition blocks in condition order:

Z = np.concatenate(
    [(P_train[i] @ B) * A[i] for i in range(n_cond) if np.any(cond_train == i)],
    axis=0,
)

but regressed it against X_train_test_genes, which stays in the training cells' natural row order. P_train[i] lists condition i's cells in their within-condition order, so concatenating yields Z sorted by condition. The two coincide only when conditions happen to be stored contiguously.

_train_cell_loadings scatters each condition's block back to the positions its cells actually occupy. Same asymptotic cost, and identical output when conditions are contiguous.

Might be worth retrying the hESC trial now to see if the R2X truly does recover when there is meaningful structure left.

---

#548 — bicv returned aggregated R2X only

BiCV rows now also carry:

column	why
Train Block R2X	In-sample R2X on the block the model was fit to
NTrainGenes, NTestGenes, NTrainCells, NTestCells	The realised block sizes
Seed	The seed determining that trial's splits and initialisation

These are columns rather than a new Metric value, so plot_bicv_r2x is unchanged. test_bicv_shape_and_range's column assertion is relaxed from an exact set to a subset, since the return shape is intentionally wider. docs/tutorial.md documents the new columns.

---

#547 — three dense blocks materialized per trial

_bicv_trial densified three blocks of the expression matrix on every trial of every rank:

block	purpose
X[train_cells][:, test_genes]	regress the held-out gene loadings
X[test_cells][:, train_genes]	project the held-out cells
X[test_cells][:, test_genes]	score the reconstruction

Each is O(cells × genes) where the matrix itself is O(nnz). Measured on a 12,000 × 4,000 dataset at 10% density the blocks alone are more than twice the data they come from. At cohort scale (1.2M × 34k, 20%/20% holdout) that is roughly 115 GB of dense allocation per trial against a 44 GB matrix.

All three are now reached through products against the full matrix, with the restriction applied by zeroing the dense operand so the result is the submatrix's without the copy:

Z^T (X[train, test] − 1 μ^T)  =  (Z_full^T X)[:, test] − (Z^T 1) μ^T

C_test now solves the normal equations rather than lstsq on the tall design, which is never formed. lstsq on the rank × rank system keeps minimum-norm behaviour when the fit is rank deficient, at some cost in conditioning.

The remaining per-trial copy is X[train_cells][:, train_genes], which run_parafac2 needs as an AnnData. Removing it would need a gene mask in parafac2.compress_dataset. We had proposed this earlier but agreed that putting cross-validation logic in parafac2 was out of scope for that repository.

[fishidaho]

@aarmey, I am not sure what experiment you had initially ran with #546 with the hESC data, but it might be worth rerunning it with the same normalization scheme you had done prior to see if the rank is no longer artificially low.