C training-IO: seeded RNG, spike encoders, IDX loader, checkpoints, reverse bridge

[w4ffl35]

Closes #27, closes #28, closes #29, closes #30 (one PR for the tight cluster; checkpoint feeds the bridge, RNG feeds the encoders). All four verified live against origin/master e1bb03d before coding: no src/data., src/rng., src/encode., src/checkpoint. exist, json.h is parse-only, and graph weights have no setter API.

What

• No seeded RNG or spike encoder in C: rate/latency/delta codings are torch-only #28: splitmix64 RNG (src/rng.) + rate/latency-linear/log/delta/random/repeat encoders (src/encode.) ported from SpikeEncoder/spikegen semantics, including torch.round half-to-even and the Bernoulli u < p equivalence contract (documented in encode.h; torch's stream is MT19937, ours splitmix64).
• No torch-free dataset loader: training data path is torchvision + DataLoader #27: IDX MNIST reader (src/data.*) with v/255 normalization matching the torch transform (ToTensor + Normalize((0,),(1)) is exactly v/255).
• No checkpoint save/load in C: checkpoints are torch pickles and C JSON is parse-only #29: JSON writer + version-stamped spikeforge-c-checkpoint v1 save/load + weight get/set accessors (src/checkpoint.*). Unknown versions, missing nodes, kind/shape mismatches, and non-finite weights fail with named errors.
• Reverse bridge missing: C-trained weights cannot flow back into NIR graphs or torch modules #30 (C side): nir_graph_export_summary re-emits a loaded graph as graph_summary-shaped JSON with current weights; tools/patch_nir_weights.py merges a checkpoint into a Python-side summary.

Only new files plus 8 lines of CMake/test registration. No edits to existing src/*.c.

Tests

34/34 ctest pass (new: test_rng, test_encode, test_data, test_checkpoint); ASan/UBSan build also 34/34. Goldens are stdlib-only generators (tools/encode_golden.py, tools/make_mnist_fixture.py) with committed fixtures, so just test needs no Python. E2E checked: C perturb -> checkpoint -> patch script output == C export bit-exact -> reloads in nir_graph_load.

Follow-ups (not in this PR)

• BLOCKED on master does not compile: two set_error call sites pass a stale third argument #36: master does not compile (two stale set_error call sites in src/graph.c, owned by another worker). Verified locally with that 2-line fix applied and reverted; this branch contains no src/*.c edits.
• Torch cross-checks need a torch env (unavailable here): SpikeEncoder parity run for the ported codings, and the spikeforge-side state_dict loader for C checkpoints (companion change in the spikeforge repo).