models : fix n_text_ctx taken from max_length instead of max_target_positions

[axiom-of-choice]

Closes #3986.

n_text_ctx sizes the decoder positional embedding (src/whisper.cpp:1797), so it must be
max_target_positions, the row count of the decoder.positional_embedding tensor this script
writes. It was reading max_length, a generation default. The audio side of the header already
uses max_source_positions, and convert-h5-to-coreml.py already reads max_target_positions.

Both fields are 448 on OpenAI's checkpoints, so this is invisible until a fine-tune sets
max_length.

Measured with a synthetic checkpoint (max_target_positions=1024, max_length=448) and
whisper-cli built from this tree:

	master	this change
fine-tune case	decoder.positional_embedding has wrong size, load fails	loads
OpenAI-style checkpoint	loads	output .bin byte-identical

#2477 and #2840 used max_target_positions as a fallback for a missing or non-int max_length;
this makes it the source, so both are subsumed. Their coercion is dropped as unreachable:
from_pretrained rejects a non-int, negative or weights-disagreeing value before the header is
written.

Not verified against a real fine-tune. #2778 may be the same bug but has no config.json to
confirm it.