Fix floating-point constant precision in compiled kernels

[Ryan11c]

Fixes #4503

Captured floating-point constants in compiled kernels are currently printed using digits10 + 1 digits. For some values this isn't enough precision to round-trip back to the same float which can make the compiled result slightly different from eager execution.

Changed print_float_constant to use max_digits10 instead. This gives enough precision for the value to round-trip correctly when the generated kernel source is compiled.

Also added tests with float32 constants that were affected by this issue. This includes 1 / 3, 128**-0.5, and 0.7071067811865476. I also updated the float64 test to use math.nextafter(1.0, 2.0) so it catches the same issue for doubles. The tests use mx.array_equal to make sure the eager and compiled results match.

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: I used GPT 5.6 Sol to help me understand the repo and how the issue is present in this repo. Fixed the bug myself and used agent to review/clean up.