fix and enable fp32 multigrid-v2 collo kernels

[Efithor]

for LDA/GGA FP32 instantiations of multigrid-v2 collocation kernels the per-grid-point pair for float the corner value underflows while the recursion start factors overflow, giving 0 * inf = NaN. Fixed via evaluating exponent directly instead of recursion product + these codepaths . Double routes and MGGA drivers are unaffected

[sunqm]

We don't plan to enable fp32 by default, as its numerical behavior is not guaranteed. You can add a cmake option to enable fp32 for users who want to experiment it in local build.

[Efithor]

We don't plan to enable fp32 by default, as its numerical behavior is not guaranteed. You can add a cmake option to enable fp32 for users who want to experiment it in local build.

@sunqm Done!