ENH: dsti-regadam model.

[ntustison]

ENH: Port syntx's RegAdam optimizer as a new dsti_regadam arm.

Adds optimizer="reg_adam" to antstorch.syn.syn_registration(): a
lightweight port of syntx's greedy.py Adam-momentum pattern (per-voxel
first/second moments, reset each pyramid level, regularizer applied to
the bias-corrected quotient instead of the raw gradient). Default
behavior (optimizer="gradient_descent") is unchanged.

Wires this into antstorch.benchmark.evaluate_mindboggle_pair() as a new
"dsti_regadam" model: same dsti regularizer and dense-SyN loop as
dsti_syn, but with reg_adam instead of plain gradient descent. dsti_syn
itself is untouched, preserving reproducibility of prior comparison
runs.

Scope: RegAdam optimizer only, not the full syntx.tvf time-varying-
velocity-field engine (~2450 lines, architecturally distinct) --
narrowed from the original request after estimating that port's size,
to test first whether the optimizer alone explains part of the dsti_syn
vs syntx dsti gap.

Also:

• tools/benchmarks/compare_syntx_antstorch_mindboggle.py: add
  dsti_regadam to --models (compared against the same syntx dsti arm
  as dsti_syn; conservative_smooth matched under --matched, optimizer
  intentionally left unmatched).
• tools/benchmarks/README.md: new model catalog for tools/benchmarks/
  (which script compares what, every antstorch model name and its
  regularizer/optimizer, syntx/antstorch/FireANTs name mapping).

Tests: 6 new tests in tests/syn/test_syn.py (unknown optimizer rejected,
loss decreases under reg_adam, default stays gradient_descent, update
trajectory differs from plain gradient descent, Adam moments reset each
pyramid level, works in 3D), 1 new test + 1 parametrization entry in
tests/benchmark/test_evaluate.py (dsti_regadam vs dsti_syn differ on
the same pair). 141/141 passing, no regressions.

Not yet run on real Mindboggle-101 data -- only validated against the
synthetic mock dataset used by the test suite.