feat(losses): add optional fused_lncc LNCC backend (loss_type='fused_lncc')

[minsuk00]

Add an optional fused_lncc LNCC backend

Wires up fused_lncc as an optional LNCC backend,
selectable via loss_type='fused_lncc'. It is a standalone fused rectangular-LNCC CUDA kernel;
this PR plugs it in as an opt-in dependency that falls back to cc when not installed, so existing
behavior is unchanged.

Changes

• fireants/losses/fused_lncc_backend.py — FusedLNCCLoss, a thin adapter matching the
  FusedLocalNormalizedCrossCorrelationLoss interface (constructor args, forward, and the
  multiscale hooks). Calls the kernel; no new math.
• fireants/registration/abstract.py — a loss_type == 'fused_lncc' branch mirroring the
  fusedcc branch (optional import + fallback to cc).
• pyproject.toml — declares fused_lncc as an optional dependency. (As a CUDA extension it must be
  installed with pip install fused_lncc --no-build-isolation against a matching PyTorch; see its README.)
• tests/test_fused_lncc_backend.py — auto-skips without CUDA or the package.

Behavior

Returns -mean(ncc), the same convention as fusedcc. Uses the exact gradient (matches
FusedLocalNormalizedCrossCorrelationLoss(use_ants_gradient=False)), rescaled so the same optimizer
hyperparameters converge identically. Scope is deliberately narrow: rectangular only, k ∈ {3,5,7,9},
mean reduction, pred-only gradient, single GPU. Other configurations (gaussian, masking, sum/none
reduction, symmetric/SyN gradients, grid-parallel sharding) raise a clear error pointing to fusedcc.
Symmetric (dual-image / SyN) gradient support is a straightforward follow-up: it runs the kernel a
second time with pred and target swapped (no CUDA change), at roughly 2x the backward cost. Left out
of this PR for now.

Performance (A40)

The loss kernel is ~3.4× faster / ~3× lighter than fusedcc. End-to-end registration is ~1.1–1.3×
per iteration (the loss is ~38% of a fused step) and ~2.7× against the non-fused cc path, at equal
VRAM and equal registration quality.

Tests

pytest tests/test_fused_lncc_backend.py passes on an A40: forward and gradient parity vs fusedcc
(k=3/5/7/9), sign/range, batched gradient scale, gradient routing, scope guards, multiscale hooks,
and an end-to-end registration through the dispatcher.