canonicalizating placeholders should ignore their `bound_var`

[lcnr]

Given an instantiated binder for<'a, 'b, 'c>, the goals

• T: Trait<'!a, '!b>
• T: Trait<'!b, '!c>
• T: Trait<'!b, '!a>

should all result in the same canonical input. They currently do not, as we simply reuse the bound_var of the placeholder in the caller as the bound_var of the placeholder in the canonical. This only impacts perf and at least for now, is likely not too relevant.

Concretely, CanonicalVarKind should only store the universe of placeholder vars and instantiating a canonical should be responsible for giving each placeholder a unique bound var.