Zeros-init + well-conditioned Gram-Schmidt; add dace_canonicalize_{cpu,gpu} frameworks

[ThrudPrimrose]

Three changes.

1. np.empty / np.empty_like -> np.zeros / np.zeros_like

256 sites across 132 files, all backends (numpy, dace, numba, cupy, dpnp, pythran, legate, jax).

An uninitialized empty buffer that a kernel writes only partially leaves stale/garbage memory in the result. Those kernels are therefore read-nondeterministic: the same input can yield different outputs on different runs, and the comparison against the reference flakes. Zero-initializing makes every run reproducible. For buffers that are fully overwritten before use the change is equivalent, so all sites are converted uniformly. Commented-out code is left untouched.

2. Well-conditioned Gram-Schmidt input

gramschmidt's initialize now builds a deterministically full-column-rank, well-conditioned matrix via a diagonal-dominance term, replacing the reject-sampling while np.linalg.matrix_rank(A) < N loop:

A = rng.random((M, N), dtype=datatype)
A[:N, :N] += N * np.eye(N, dtype=datatype)

A plain random matrix is only full-rank in expectation and can still be poorly conditioned, which makes the QR numerically unstable and sensitive to harmless FMA-contraction reassociation. The added diagonal gives cond(A) ~= 1.5 deterministically. All framework implementations validate against the same-input NumPy reference, so the change is self-consistent.

3. New frameworks: dace_canonicalize_cpu / dace_canonicalize_gpu

Two frameworks that run DaCe's canonicalize pipeline instead of auto_optimize, to contrast the WCR (write-conflict-resolution) reduction lowering:

• The new frameworks build the SDFG, run canonicalize(target=cpu|gpu, peel_limit=4, break_anti_dependence=True, interchange_carry_with_map=True, scatter_to_guarded_maps=True) + finalize_for_target, leaving sdfg.openmp_array_reductions = True (WCR-config ON) so a whole-buffer WCR accumulator of a parallel map lowers to an OpenMP reduction(op:A[0:n]) array-section clause instead of per-element atomics.
• The existing dace_cpu / dace_gpu now set openmp_array_reductions = False explicitly before compile (WCR-config OFF, the default per-element atomic path).

Registered exactly like dace_cpu/dace_gpu: framework_info/dace_canonicalize_{cpu,gpu}.json + a DaceCanonicalizeFramework class (a DaceFramework subclass shared by both arches, selected by the JSON arch field). Reuses the existing *_dace.py implementations and every DaceFramework calling convention.

Verified: the CPU variant builds, runs and validates against NumPy on atax (a reduction kernel). The GPU variant is defined/registered analogously and builds + canonicalizes; fully running it needs the GPU device-schedule assignment that DaCe's canonicalize(target='gpu') path does not yet emit, so it is definition-only for now.