fix: improve model_args type coercion in handle_arg_string

[ManasVardhan]

Summary

Improves the type inference/coercion logic in handle_arg_string() (lm_eval/utils.py) to fix several edge cases where model argument values end up as the wrong Python type.

Fixes #2183 (related: #2167)

Changes

handle_arg_string improvements:

• Negative integers: -1 now correctly becomes int(-1) instead of staying a string (the old str.isnumeric() check only matched non-negative digits)
• None detection: None/none now returns Python None
• Explicit quoting: wrapping a value in quotes forces it to stay a string, e.g. revision="123123" keeps the value as str instead of casting to int. This is the recommended fix for the revision issue in [BUG] Huggingface and Neuron runs fail if model revision is an integer #2167.
• Scientific notation: 1e-5 correctly becomes float(1e-5) via the float fallback

Tests added:

• Comprehensive tests for handle_arg_string covering bools, None, ints, negative ints, floats, scientific notation, plain strings, and explicit quoting
• Tests for simple_parse_args_string covering end-to-end parsing with type coercion

Backward compatibility

All existing behavior is preserved. The only additions are:

1. Negative integers and None are now properly coerced (previously returned as strings)
2. Users can now quote values to opt out of coercion (new feature, no existing behavior changed)

[baberabb]

Thanks for the PR! LGTM