Make the vision tower a knob instead of a flag-order accident

--language-model-only was hardcoded in both exec lines, so the only way to
serve images was to pass --no-language-model-only through EXTRA_ARGS and rely
on argparse taking the last flag. That breaks silently when the exec line is
edited: images are still accepted and still counted as prompt tokens, and the
model answers from placeholder embeddings.

VISION=0 stays the default and the command line is unchanged. VISION=1 swaps in
--limit-mm-per-prompt and --mm-processor-kwargs; both stay overridable from
EXTRA_ARGS, which is expanded after them, so only the flag that cannot be
countered gets a knob.

Gotcha 9's 2.7 GB for the tower does not hold on these checkpoints. model.visual.*
sums to 0.858 GiB of BF16 in both the base dir and the -fast variant, and a
runtime A/B agrees: model loading 15.13 GiB against 14.26, same server, same
config, with non-weight overhead 0.42 against 0.41 GiB either way. The tower is
BF16 in both dirs, so quantization is not the explanation; where 2.7 GB came
from I could not work out.