Fixes a bug in serializing min/max plus one more. (#35850)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35850
1. Clamping values were not being propagated through all the structures
and hence were not being serialized.
2. Moved to using Scalar for min/max instead of float. Reason being, the
fusion for hardtanh_ does not work. During sub graph rewrite we direct
values from hardtanh_ to preacking ops, but since they expect float
values, the types conflict and we cannot serialize the model.
Test Plan: Imported from OSS
Differential Revision: D20807523
fbshipit-source-id: 57d6b2e4b65afd9510a0f3ba9365333b768977f5