allow per-language discriminants

[bjchambers]

For languages like SymbolLang that don't have an outer-most enum this allows defining a discriminant and reducing the set of nodes considered for matching. This allows each Language to determine the type to use for the discriminant and how to compute it.

For instance:

1. This may be a std::mem::Discriminant<Self> if the type is an enum
2. This may be a u64 if it is easy to hash the "operation"
3. This may be Symbol for something like SymbolLang (or a discriminant of the operation, in the case of something like SymbolLang but using an enum for the symbol)
4. This may be something a bit richer for cases like ENodeOrVar.

This also allows languages to trade-off how rich the discriminant is with how efficient it is to compute / compare.

[mwillsey]

This looks great! Could you please document the associated type as well as in the define_language macro that it's using the enum discriminant.

[bjchambers]

This looks great! Could you please document the associated type as well as in the define_language macro that it's using the enum discriminant.

Absolutely. Wanted to get it out and make sure it seemed reasonable before polishing it too much. Will do so now!

[bjchambers]

I wasn't sure about this check here. Why is it expected that nodes with equal discriminant will be adjacent? It doesn't seem like the nodes are sorted by discriminant (the Discriminant doesn't implement Ord).

[bjchambers]

Comments added.