fix: call .eval() on tokenizer and model in KronosPredictor

[hexonal]

KronosPredictor never switches the modules out of training mode, so inference runs with dropout active.

Impact

Kronos-small ships attn_dropout_p = 0.1. Because KronosPredictor.__init__ only does .to(device), the modules stay in train() mode for the whole lifetime of the predictor:

• On MPS it crashes outright. F.scaled_dot_product_attention with a non-zero dropout_p is not implemented for the MPS backend, so any predict() on Apple Silicon raises rather than returning a forecast.
• On CUDA/CPU it is worse than a crash — it is silent. Predictions come back looking fine, but every forward pass drops 10% of the attention weights at random. Two calls with identical inputs give different answers for a reason that has nothing to do with the models sampling behaviour, and the extra variance is indistinguishable from the intended sample_count sampling noise.

Measured on this fork: with sample_count=5 on 60m US equity bars, adding .eval() cut MAE by roughly a fifth on an otherwise unchanged pipeline.

Fix

self.tokenizer = self.tokenizer.to(self.device).eval()
self.model = self.model.to(self.device).eval()

.eval() is the conventional place for this — the predictor is inference-only, and nothing in the class ever calls .train(). Callers who genuinely want stochastic modules can still flip them back after construction.

Two lines, no behavioural change for anyone who was already calling .eval() themselves.