Bump sgd_step_batch16 micro pins again (55 → 100 µs): runner noise

Same benches flaked again at 63/74 µs vs the 55 µs pin — up from 40/38 µs
the previous run, i.e. ~2x run-to-run variance on identical code (yscv-
model + bench chain still byte-identical to f364661; criterion still
reports no change, p > 0.05). The short batch16 steps (~true 20 µs) are
noise-dominated on the shared runner while the longer batch64 (208 µs)
stays stable. 100 µs = worst observed (74) × the ~1.35 margin used for
the batch64 pin. These gates only catch gross regressions here anyway.

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=55
+model_train_step_modes/sgd_step_batch16.max_upper_us=100
 model_train_step_modes/sgd_step_batch64.max_upper_us=280
-model_train_step_modes/sgd_step_batch16_hinge.max_upper_us=55
+model_train_step_modes/sgd_step_batch16_hinge.max_upper_us=100