[P4Testgen] Revisit handling of varbit in P4Tools/P4Testgen

[vlstill]

The current handling of varbit fields in testgen is somewhat cumbersome and leads to some non-well-typed expressions. When the program contains varbit field, the testgen needs to remember the actual extracted size and it needs to work with the field accordingly.

• When the field is emitted, the actual size is emitted.
• When one varbit field is assigned into another (possibly as part of a header), the size of the source is also propagated to the destination. This may change the actual size of the destination.

The current solution uses SizedVarbit/Extracted_Varbit type, which has some problems. E.g., when there is just a varbit extracted (e.g. testgen-p4c-bmv2-protobuf-ir/issue447-bmv2.p4), the packet expression is something like 0 ++ (SizedVarbit<32><16>)|pktvar_0(bit<16>)| which is of type bit<16> but the optimizer turns it into (SizedVarbit<32><16>)|pktvar_0(bit<16>)| which is SizedVarbit<32><16>.

Looking at the other code that relates to varbits, it seems like these are mostly special cases that treat SizedVarbit<M><N> as bit<N>. However, because of varbit assignments, it is not possible to just replace varbit<M> with bit<N> when it is extracted, as that would lead to wrong assignments. At the same time, it is not sufficient to keep the value just as varbit<M> in the header as ween need to keep the runtime size known.

One possibility I can think of would be to save bit<N> in the packet buffer, but something like VarbitCast<M, N> in the header's varbit field. VarbitCast<M, N> would be an expression that would contain a bit<N> typed expression, but it would itself be of type varbit<M>. This way, it would be possible to assign these expressions and the program should be well typed.

More discussion on the problem can be found in PR #4300.