mixqtl branch: run on current NumPy/PyTorch and match mixqtl.R nominal output

[fbyukgo]

We've been using the mixqtl branch on our data (~250 donors) and made a small set of fixes so it runs on current NumPy/PyTorch and reproduces mixqtl.R's nominal output. Sharing them back in case they're useful.

What's here:

Dependency updates (P1–P2): NumPy 2.0 (np.NaN→np.nan, np.float→float) and PyTorch 2.0 (torch.solve/qr/symeig → their torch.linalg equivalents). No behaviour change — just needed to import/run.
mixqtl.R parity in map_nominal/trc_calc (P4–P7, these change results):
NaN-initialise the meta/asc arrays — ASE-absent genes were emitting uninitialised memory as beta_meta (a bug).
Meta rule follows rlib_meta.R: trc-fallback when ASE is absent, and the n_cutoff gate (branch had neither).
trc_calc fits the covariate offset on all count-observed individuals before the count filter (as R does), instead of re-fitting on the survivor subset — the old order is unstable/biased on low-expression genes. (Also fixes an SE bug where the phenotype shape gave dof = -1.)
p-value convention matched to R (z above n_cutoff, else t with df = n).

Two of these (NaN-init, and the covariate order) are genuine correctness fixes, not just cosmetic alignment.

Validation. With trc_cutoff matched to R, beta_trc/beta_asc/beta_meta agree to ~4 dp (r ≈ 0.9997, ~0.02% sign flips, scale ratios ~1.0, p-values match after P7), and it runs ~100× faster than the R split job. The concordance script we used is included as a regression check.

One note: count_threshold is a call-site literal in map_nominal; might be worth exposing as a parameter (we left the existing default untouched).

Happy to split this differently, and — the branch is nominal-only, so if there's interest we'd be glad to contribute our permutation-based eGene layer for the ASE-integrated statistic as a separate PR.

Full patch descriptions:

P1 (NumPy 2): np.NaN→np.nan (genotypeio.py, mixqtl.py, eigenmt.py); np.float(→float( (eigenmt.py).
P2 (PyTorch 2): torch.solve(B,A)→torch.linalg.solve(A,B); torch.qr→torch.linalg.qr; torch.symeig(…,eigenvectors=False)→torch.linalg.eigvalsh (mixqtl.py, core.py, eigenmt.py). Equivalent numerics.
P4 (bug): meta/asc arrays were np.empty and only written when ASE existed → ASE-absent genes emitted uninitialised memory as beta_meta. Now np.full(…, nan) (and -1 for int columns).
P5 (meta rule = rlib_meta.R::my_meta_fast_): start from the larger-n arm; per-SNP fall back to the other arm where the primary is NaN; inverse-variance combine only when both arms have ≥ n_cutoff (15) samples; when ASE is absent, meta = trc. Branch previously combined only when ASE existed, with no n-gate and no fallback.
P6 (bug / sounder estimator): fit the covariate offset on all count-observed individuals before the count filter, subtract it, then drop low-count individuals, then regress on genotype with no covariate re-fit (as mixqtl.R). Old order re-fit covariates on the survivor subset → rank-deficient when survivors < #covariates and biased (survivors selected on expression). Also: pass phenotype 1-D so dof = n-2 (was giving -1 → NaN SE).
P7 (p-value = get_pval_fast_): z when the relevant n > n_cutoff, else Student-t with df = n; meta's effective n tracked per row. Closes the last -log10 p gap on small-n genes.