stat_tracking: fix AttributeError on second update() call caused by in-place np.stack overwrite

[Dev-X25874]

What

PerPromptStatTracker.update() crashes with AttributeError: 'numpy.ndarray' object has no attribute 'extend' on any call after the first when the same
prompt key is seen again.

Root Cause

The second for loop in update() overwrites self.stats[prompt] with the
result of np.stack(...), converting the stored value from a list to a
numpy.ndarray. On the next call, the first for loop tries to call
.extend() on that ndarray, which does not exist on that type.

# before — mutates self.stats[prompt] to ndarray
self.stats[prompt] = np.stack(self.stats[prompt])

Fix

Compute the stacked array into a local variable prompt_history instead of
overwriting self.stats[prompt], so the list is preserved across calls.

# after — self.stats[prompt] stays a list
prompt_history = np.array(self.stats[prompt])

All downstream mean/std calculations use prompt_history; no values change.

Impact

• Affects all training scripts that rely on cross-call reward history
  accumulation (train_sd3.py, train_flux.py, train_flux_fast.py, etc.)
• The built-in main() smoke test only calls update() once before clear(),
  so the bug was not caught locally
• One-line change, zero behaviour delta