use native tabicl inference

[borawhocodess]

another of the remaining todos in #43
follows the dropout context manager comment in #40

in this pr:

• deleted train_path_without_dropout
• explicit inference device replaces the device workaround
• native eval already disables dropout
• keeps native amp and casts outputs to float32 for numpy

not here:

• inference batching with different class counts