Finetuner performance + robustness optimizations

[r-kellerm]

1. AMP for the full fine-tune path - added use_amp/amp_dtype args plus _is_cuda_device/_amp_available/_pick_amp_dtype helpers and a _autocast() context. Mixed precision is gated to amp_active = use_amp and not only_proj_layer, so the cached linear-probe path stays fp32 (avoids the bf16 to numpy cache crash). GradScaler is wired only for fp16; bf16 is preferred when supported and needs no scaling.
2. zero_grad(set_to_none=True) - replaced plain zero_grad().
3. Backbone freeze - on the probe path, backbone params are set requires_grad_(False) once and restored in a finally, replacing the per-batch no_grad()+detach().
4. Robustness - avg_loss initialized to nan (no NameError on empty loader), and fwd_cache.clear() moved into the finally so the cache is released even on exceptions.