bug: setup interactive model selection crashes on invalid input

[weklund]

Description

_prompt_model_selection() in cli/setup.py has no input validation, causing multiple crash/correctness bugs when users enter invalid model selection strings.

Bugs Found

1. Non-numeric input crashes with ValueError

Entering abc at the model selection prompt causes an unhandled ValueError: invalid literal for int() with base 10: 'abc'.

2. Out-of-range index crashes with IndexError

Entering 99 (when only ~111 models are listed) causes IndexError: list index out of range.

3. Input "0" silently selects wrong model

Entering 0 maps to Python index -1 (last element in the list), silently selecting the wrong model.

4. Negative numbers wrap around

Entering -1 maps to index -2, silently selecting the wrong model via Python negative indexing.

5. Invalid quantization accepted without validation

Entering 1:fakeq passes fakeq as the quantization without any validation. This will fail downstream.

Expected Behavior

• Non-numeric input should show a retry prompt with a clear error message
• Out-of-range indices should show "Please enter a number between 1 and N"
• Zero and negative indices should be rejected
• Quant values should be validated against the allowed set (bf16, int4, int8)

Steps to Reproduce

# Direct test
from mlx_stack.cli.setup import _prompt_model_selection
# ... feed invalid inputs to the prompt

Or run mlx-stack setup interactively and enter abc, 99, 0, -1, or 1:fakeq at the model selection prompt.

Impact

High — this is the primary onboarding flow. Any user who makes a typo during interactive setup will see a Python traceback instead of a helpful error message.