test: add missing tests for signed/field numeric generics and fix proptest bug

[AztecBot]

Summary

Adds comprehensive test coverage for signed integer and Field arithmetic in numeric generics, which were identified as having near-zero test coverage. Also fixes a bug in the existing proptests.

Changes

Bug fix — proptest never generated negative signed values:

• In types/arithmetic.rs line 592, integral_maximum_size() was called instead of integral_minimum_size() for signed types. For i8, this meant the proptest only generated values in [127, 127] instead of [-128, 127]. All three proptests (instantiate_before_or_after_canonicalize, instantiate_before_or_after_canonicalize_checked_cast, compare_to_comptime) were affected.

New Integer unit tests (39 tests):

• Signed addition/subtraction/multiplication/division with negative values
• Overflow and underflow detection for signed types
• Signed negation (including i8::MIN overflow case)
• Signed comparison with negatives
• Type mismatch returns None
• Wider signed types (i32, i64) with negative arithmetic
• Field arithmetic (addition, subtraction wrapping, division by zero, remainder rejection, negation, comparison)
• is_negative behavior for signed, unsigned, and Field
• as_field vs as_field_twos_complement encoding of negative values

New numeric generic frontend tests (8 tests):

• Signed numeric generics in function signatures, structs, traits, where clauses
• Signed generics with i8, i64 types
• Negative value turbofish syntax (foo::<-5i32>())

New arithmetic generic frontend tests (14 tests):

• Signed: addition, subtraction producing negatives, multiplication with negatives, division with negatives, variable cancellation, constant folding, overflow detection, underflow detection
• Field: addition, subtraction, multiplication, variable cancellation, constant folding, large values (> u32::MAX)

Context

See discussion in the Slack thread — Tom French noticed that Field and signed types were allowed in numeric generics (likely as a result of the Integer refactor making everything more uniform). Jake Fecher suggested adding tests rather than restricting the types. This PR addresses the identified coverage gaps.

Previous analysis: https://gist.github.com/AztecBot/2801f88caf43211a3df05774bff9886b

ClaudeBox log: https://claudebox.work/s/c192b26735dbc607?run=2