No switch to Modular<float>/Modular<double> in DelayedTag mode for fgemm

[ClementPernet]

Currently, fgemm with DelayedTag mode tests whether it is possible to convert the arguments over floats or double (in order to compute faster). This switch is too low in the call stack and e.g. prevent the user to force the use of ModularBalanced<int64_t> when she knows what she does (e.g. benchmarketing igemm). This switch should happen earlier, and/or be integrated with the ConvertTo modes.