Lower together SFT batch size to 8 (tensorzero#3180)

Their API started rejecting 32 with
"Invalid param in Model request: batch size is higher than 8":
https://github.com/tensorzero/tensorzero/actions/runs/17057954076/job/48359010125

We should make this parameter configurable at some point

Co-authored-by: Viraj Mehta <[email protected]>

Author: Aaron1011

tensorzero-core/src/optimization/together_sft/mod.rs

@@ -394,7 +394,7 @@ impl Optimizer for TogetherSFTConfig {
                 n_checkpoints: Some(1),
                 n_evals: Some(n_evals),
                 learning_rate: 0.00001,
-                batch_size: 32,
+                batch_size: 8,
                 lr_scheduler: TogetherLRScheduler {
                     lr_scheduler_type: TogetherLRSchedulerType::Linear,
                     lr_scheduler_args: TogetherLRSchedulerArgs { min_lr_ratio: 0.0 },