Add C weight initializers, LR schedules, and extra losses

[w4ffl35]

Closes #45, closes #46, closes #47.

What

New training-side modules (headers in include/nir_c_runtime/, sources in src/, registered in both CMake files). No changes to backward ops, STDP, driver, or checkpoint-state code.

#45 — weight initializers (init.h/init.c)

• nir_init_uniform, nir_init_xavier_uniform/_normal, nir_init_he_uniform/_normal (torch kaiming fan_in mode), all drawing from nir_rng_t.
• nir_init_fans implements torch _calculate_fan_in_and_fan_out for Linear/Affine [out, in] and Conv [out, in, *kernel] shapes.
• Formulas and fan conventions documented in the header.

#46 — LR schedules (schedule.h/schedule.c)

• Stateless nir_lr_step_decay, nir_lr_cosine, nir_lr_exponential mirroring torch StepLR / CosineAnnealingLR / ExponentialLR with last_epoch = step.
• Step 0 yields base_lr; the counter is the checkpoint step (completed optimizer steps), so resume continues the curve. Header shows the per-epoch opt->lr write for both SGD and Adam.
• Cosine follows torch past t_max (rises again, no clamp) — documented.

#47 — losses (loss.h/loss.c)

• nir_mse_loss: mean-reduction MSE + gradient, mirroring F.mse_loss on one sample; serves readout/membrane regression recipes.
• nir_spike_count_loss: the decided form is cross-entropy over spike counts (not MSE on counts), matching the snnTorch sum-over-time-then-CE recipe and spikeforge's averaged-readout-plus-cross_entropy path. Caller sums the T-step window; each step's spikes share their class's count gradient.
• Both keep the single-sample-first 1-D convention; batch reduction stays with Mini-batch training (gradient accumulation, batched loss reduction, seeded shuffling) #42.

Tests

• test_init: fan values incl. Conv shapes, mean/variance within tolerance at N=20000, seeded bit-identical reproducibility, error shapes.
• test_init_graph: Xavier-init fc1 weights land via nir_graph_set_weights (read back byte-identical) and the graph runs forward cleanly.
• test_schedule: closed-form values incl. boundary steps (9/10, 19/20, T/2, T, 2T), SGD + Adam composition through lr, error shapes.
• test_loss: closed-form MSE and spike-count value + gradient, spike-count bit-equals nir_cross_entropy on the same vector, aliasing rules.
• Local: 39/39 ctest green in Debug, Release, and ASan/UBSan builds, no compiler warnings under -Wall -Wextra -Werror.