Persist optimizer state in checkpoints (closes #43)

[w4ffl35]

Checkpoints persist weights plus meta but drop SGD velocity and Adam m/v/step, so resuming silently restarts the optimizer (wrong for Adam bias correction).

This change adds an optional v2 optimizers section keyed by node name:

• SGD entries carry velocity plus optional velocity_bias; Adam entries carry step/m/v plus optional step_bias/m_bias/v_bias.
• Hyperparameters are not persisted; re-init with the same hyperparameters before load for a bit-identical resume (documented in checkpoint.h).
• v1 files still load; caller optimizers absent from the file keep their state. Plain loads/save stay weights-only and ignore the new section.
• Buffer shapes validate against the graph with named errors mirroring the weight-mismatch discipline; every file optimizer entry must match a caller entry.
• tools/patch_nir_weights.py accepts v1/v2 (weights only, optimizers ignored).

Tests: test_checkpoint gains SGD and Adam save-fresh-load-N-steps bit-identical resume checks (memcmp on weights and state, step counters asserted), v1-compat, plain-ignores-optimizers, file round-trip, and save/load reject coverage. Full ctest passes 41/41 on Debug, Release, and ASan/UBSan locally.