Dispatch interpolate_ms_features to the fused quantem-cuda kernel

[cedriclim1]

What

interpolate_ms_features (non-tilted KPlanes) now dispatches to quantem.cuda.core.ml.kplanes_fuse — the fused fwd+bwd CUDA kernel added in cedriclim1/quantem-cuda#1 — when the inputs qualify (CUDA, fp32, standard [B,3] points and [3,C,H,W] grids). All other cases (CPU, fp64, missing quantem-cuda) fall through to the unchanged torch path; the import is guarded so installations without quantem-cuda see zero difference.

Why

F.grid_sample backward is the dominant cost of non-tilted KPlanes INR training. The fused kernel benchmarks at ~9.6x fwd+bwd on the isolated op; end-to-end this translates to:

200³ phantom, 29 tilts, batch 8192, RTX PRO 6000 (idle GPU pinned)	s/epoch
torch path (two runs)	2.245 / 2.267
fused dispatch (two runs)	1.481 / 1.601

~1.45x end-to-end per epoch, with final losses in family (1.04–1.08e-3 after 4 epochs on both paths).

Dependencies / conflicts

• Depends on Add fused non-tilted K-Planes interpolation kernel (kplanes_fuse) quantem-cuda#1 at runtime for the fast path (benchmarked against a wheel built from that branch). Without it the guarded import leaves the torch path in charge, so merging this first is safe — it just does nothing until quantem-cuda ships the op.
• Trivial textual conflict with Hoist loop-invariant sampling grid in CP-TILTED interpolation #13, which touches the coordinate-building lines of the same function; whichever merges second needs a mechanical rebase.

Evidence

• New TestFusedDispatchParity in tests/ml/test_kplanes.py: fused-vs-torch forward parity (rtol 1e-4) and gradient parity for points + all grids (kernel toggled via monkeypatch), fp64 fallback, CPU path; skipped automatically when CUDA or quantem-cuda is absent. 6/6 pass with the kernel installed.
• Full suite uv run pytest tests --runslow: 433 passed, 1 skipped; only the pre-existing unrelated ptychography failure.
• ruff check / format clean on touched files.

Side finding (not addressed here)

KPlanes(use_hybrid_mlp=False) — the constructor default — crashes in get_params() with AttributeError: 'KPlanes' object has no attribute 'sigma_net' because get_params/param_keys reference sigma_net unconditionally while __init__ only creates it under use_hybrid_mlp=True. Hit while benchmarking; flagged for a separate fix.

[cedriclim1]

Closing without merge: this dispatches to quantem.cuda.core.ml.kplanes_fuse, which quantem-cuda does not export (inventory: kplanes_tilted_fuse, kplanes_tilted_tv_fuse, tv_loss_{iso,sq,l1}_3d). The guard makes it a permanent no-op — the ImportError sets the handle to None and every call takes the torch fallback — and it also bypasses the use_cuda_kernels config policy that #21 standardizes. Revisit if quantem-cuda grows a non-tilted kernel; until then #13's hoist is the live optimization for this path.