Bump sgd_step_batch16 micro pins for slow CI runner (26 → 55 µs)

The model_runtime_ops gate failed on sgd_step_batch16 (40.2 µs) and
sgd_step_batch16_hinge (38.6 µs) against a 26 µs pin. Not a regression:
yscv-model and its whole bench dependency chain (optim/autograd/tensor/
kernels) are byte-identical to the last merge (f364661) — the only
commits touching them, e3cffe0 and its revert d0b26e4, cancel out — and
criterion itself reports "No change in performance detected" (p > 0.05).
The Actions runner is simply slower/noisier; same reason as ac1d3ef and
bc6dad4 bumped the sibling pins. 55 µs keeps ~35% headroom over observed.

Author: enthropy7

benchmarks/ci-baseline-model-micro.txt

@@ -5,6 +5,6 @@
 
 model_forward_modes/linear_64x32_batch32.max_upper_us=7
 model_forward_modes/linear_relu_linear_batch32.max_upper_us=16
-model_train_step_modes/sgd_step_batch16.max_upper_us=26
+model_train_step_modes/sgd_step_batch16.max_upper_us=55
 model_train_step_modes/sgd_step_batch64.max_upper_us=280
-model_train_step_modes/sgd_step_batch16_hinge.max_upper_us=26
+model_train_step_modes/sgd_step_batch16_hinge.max_upper_us=55