WIP: refactored KK.

[ntustison]

ENH: Add shared RegAdam infrastructure for DiReCT

• add dedicated direct and registration namespaces
• extract reusable RegAdam state and direction updates
• update SyN and RegAdam to use the shared implementation
• fix Gaussian fallback for DST-I and unregularized modes
• document the planned DiReCT module organization
• add RegAdam state and dispatch tests

ENH: Share RegAdam via antstorch.registration; scaffold antstorch.direct.

• Extract the Adam-moment step-direction math (duplicated between
  RegAdam.step() and _fit_syn_level()) into reg_adam_direction()/
  RegAdamState in a new antstorch.registration package, so higher-level
  models (SyN, and eventually DiReCT) share one implementation instead of
  depending on each other's private modules.
• Fix RegAdam.step()'s regularizer dispatch: setting gaussian_sigma no
  longer silently falls through to Gaussian smoothing when regularizer is
  'dsti'/'dsti1'/'none' (old condition only excluded 'sobolev'); now
  requires regularizer == 'gaussian' explicitly. Covered by a new
  regression test that fails if Gaussian smoothing is invoked for those
  regularizers.
• Scaffold antstorch.direct (DiReCT cortical-thickness tensor port, no
  implementation yet) with a README defining module boundaries and two
  reuse rules for whoever writes core.py/regularization.py next:
  regularization.py must share antstorch.syn.core.smoothing rather than
  reimplement Gaussian/Sobolev/DST-I/B-spline filtering, and
  antstorch.direct must never be re-exported as antstorch.cortical_thickness
  / antstorch.longitudinal_cortical_thickness, since those names are
  already public API in antstorch.utilities.cortical_thickness.