Fix: scale ray infeasibility back to original units

[LucasBoTang]

Both infeasibility tests compare a ray's residual against that ray's objective value. The objective is converted back to the original problem's units; the residual is not.

Fix

1. src/utils.cu: divide the two ray infeasibilities by constraint_bound_rescaling / objective_vector_rescaling (dual_slack_norm stays raw — it only normalizes the ray and cancels).
2. src/utils.cu: eps_infeasible default 1e-12 → 1e-10. The fix raises every ratio, so the old constant is over-tight.
3. src/cli.c: add --eps_infeasible <tol>.

Results

Mittelmann-49 + MIPLIB-383 feasible, Netlib-infeas-29 + MIPLIB-infeas-334, all with the fix applied.

eps_infeasible	Netlib	MIPLIB-infeas	false positives
1e-6	27/29	312/334	7
1e-7	26/29	310/334	2
1e-8	25/29	310/334	1 (leo2)
1e-9	25/29	305/334	0
1e-10	24/29	300/334	0
1e-11	23/29	284/334	0
1e-12	21/29	265/334	0

The lowest ratio any feasible instance reaches is 5.176e-09 (leo2). 1e-10 is the first clean value and keeps the 21–28× margin the tolerance was originally calibrated to.