Fixes #149794

Summary

The previous implementation converted the entire Duration to f64 before multiplying/dividing, which caused precision loss because Duration has ~93 bits of precision (u64 secs + u32 nanos) but f64 only has 53 bits of mantissa.

This caused issues like:

• Duration::MAX.mul_f32(1.0) panicking due to overflow
• d.mul_f64(1.0) not preserving the exact value

Solution

Multiply/divide seconds and nanoseconds separately, then combine the results. This preserves precision for the common case while properly handling edge cases.

Changes

• Add helper function duration_from_float_secs() to reduce code duplication
• Add #[track_caller] to all four methods for better panic messages
• Add explicit infinity check with is_infinite()
• Check overflow before casting to u64 and use checked_add()
• Clamp final nanoseconds with .max(0.0) to handle floating-point rounding errors
• Update doctest expected values to match new (correct) behavior
• Add comprehensive tests for edge cases (overflow, infinity, subnormals, zero, boundaries)

Test plan

• All existing time tests pass (70 tests)
• All doctests pass
• New tests cover: overflow, infinity, negative infinity, subnormals, zero duration, rounding boundaries, small durations, large factors