[antonio-dimeglio]

Adds NODDI to Larmor: larmor.dwi.NoddiModel / NoddiFit and the lar noddi command, with a guide page.

Method. Each voxel starts from a dictionary of NODDI signals about its DTI direction, solved as AMICO
does (free water by NNLS, the tissue support by a non-negative L1 + L2 solve, then NNLS on that support).
From there it is fitted as the NODDI Matlab Toolbox 1.05 fits WatsonSHStickTortIsoV_B0: the toolbox's
order-12 series for the Watson sticks and its Watson-averaged zeppelin, S0 from the positive b0s, sigma
from their spread divided by 100, the Rician likelihood, the toolbox's parameter encoding and its two
stages (fractions first, then all five parameters including the direction). The optimiser is a bounded
BFGS with numerical gradients in place of MATLAB's fmincon. The whole voxel loop is C++ on all threads.

Supporting changes.

• solve_ridge_nnls takes an L1 weight (with x >= 0 it is a linear term, so it shifts the
  correlations, and the ridge rows of the QR path get -l1 / ridge); bound as _core.ridge_l1_nnls.
• The fits read the masked voxels in chunks (signal_rows, signal_chunks, finite_rows in
  larmor.dwi.inputs) instead of copying every brain voxel as float64 first: DTI, DKI, the WM response,
  CSD and its response, MSMT-CSD and its responses, and mFOD (which also turns each chunk's amplitudes
  into SH before the next). At the default chunk size the outputs are identical to before on clean and
  damaged data; with chunks of a few voxels they differ only by rounding (about 1e-8 relative in
  MSMT-CSD), which tests/dwi/test_chunked_reading.py now checks.
• compartment_attenuation releases the GIL; the scaled Bessel I0 moves to common/maths.

Validation (reference data in tests/data/noddi, made by make_noddi_phantom.py and
tests/matlab_scripts/make_noddi_ref.m):

• Noise-free phantom: the fit recovers the truth to < 1e-3 (measured 5e-5); the toolbox to 3e-4.
• Noisy phantom (SNR 30): errors equal the toolbox's (v_ic 0.037 for both); voxel by voxel the mean
  difference to the toolbox is 0.0007; the 2 voxels that differ by more than 0.02 have a lower negative
  log-likelihood with Larmor.
• HCP subject 100307, a 1104-voxel crop: mean difference to the toolbox 0.000 (v_ic), 0.001 (OD),
  0.000 (v_iso), directions within 0.0 degrees median; 0.3 s against the toolbox's 211 s.
• Whole HCP brain (741k voxels, 288 volumes): 3.5 min, peak memory 3.4 GB.
• C++ tests check the series against the exact Watson integral, the whole signal against Zhang et al.'s
  equations, the coefficient joins, the averaged zeppelin, and noise-free recovery including the direction.