Fix Unsloth autocast dtype for bf16 models

[vivekkalyan]

Summary

PipelineTrainer + LocalBackend can load bf16 base models like Llama 3.1 in 16-bit mode while ACCELERATE_MIXED_PRECISION is unset. We were defaulting the Unsloth logprob path to fp16 in that case, which can hit Half vs BFloat16 matmul mismatches.

This keeps explicit mixed-precision settings authoritative, but when the env var is unset, infers the autocast dtype from the loaded model parameters. Unknown mixed-precision/model dtype states fail early instead of silently falling through.

Validation

• .venv/bin/pytest tests/unit/test_unsloth_autocast_dtype.py -q
• Sky 2x H200 smoke: Llama 3.1 + PipelineTrainer + dedicated LocalBackend, load_in_4bit=False, load_in_16bit=True, ACCELERATE_MIXED_PRECISION unset; reached step 2 and reloaded adapters for steps 1 and 2.

[arcticfly]

Thanks @vivekkalyan!