Fix numerical abs usage in primal-dual residual gap computation

[LucasBoTang]

Summary

This PR fixes an incorrect use of abs when computing the log-scaled primal-dual residual gap during adaptive restart, as noted in issue #60.

The previous implementation could invoke the integer version of abs, leading to unintended type conversion and incorrect behavior for floating-point values.

Changes

• Replace abs(...) with fabs(...) when computing primal_dual_residual_gap in perform_restart()

[LucasBoTang]

Hi @ZedongPeng,

I ran a benchmark comparison between the current version (abs) and the test version (fabs).

From the plot, the solve time is essentially unchanged.

The iteration counts differ slightly, and in two cases the difference exceeds 1%.

[WARNING] more than 1% mismatch in Iterations for instance ns1687037:
             Status Iterations Primal objective Dual objective
Original TIME_LIMIT  246429000      4.678306582    -250.390681
Test      TIME_LIMIT  252143200      4.689183513   -249.5551493
--------------------------------------------------------------------------------------------------

[WARNING] more than 1% mismatch in Iterations for instance neos-4535459-waipa:
             Status Iterations Primal objective    Dual objective
Original  TIME_LIMIT    7857200   3.06017374e+15  -4.356248758e+10
Test     TIME_LIMIT    7895200  3.056019324e+15  -4.372484483e+10
--------------------------------------------------------------------------------------------------