Skip bytecode inference of the actual value for `assume`.

To do so, I've generalized our existing support for skipping bytecode inference, which I had originally set up for `ExpectFailure` in https://github.com/google/truth/commit/3caa0e845575c2c6f17ee62716f7064584b79efb + https://github.com/google/truth/commit/24b5a310fd899b5554ded8287f33c826f5d7778b.

This change improves the performance of failed assumptions. It makes the messages of the resulting `AssumptionViolatedException` somewhat worse, but my sense is that no one really looks at those. As discussed in b/478281659, we could go further by also avoiding cleaning stack traces, which probably also no one looks at.

RELNOTES=n/a
PiperOrigin-RevId: 860538616