Fix slow compilation of programs using recursive union types

[SeanTAllen]

The subtype checker's assumption stack (which breaks cycles during recursive type checks) only operated on nominal types. When a recursive union type appeared as a field in its own members, the compiler would re-expand the full union membership at every level of recursion, causing exponential work in the expression pass.

Three changes fix this:

1. Generalize exact_nominal to exact_type, which can structurally compare any type AST node (unions, intersections, tuples, arrows, type params, capabilities, ephemeral markers) — not just nominals. This lets the assumption stack recognize compound type pairs. Importantly, exact_type compares nominal typeargs structurally rather than via is_eq_typeargs (which calls is_eqtype and re-enters the subtype checker, causing infinite recursion when called from check_assume during union subtype checks).

2. Add assumption stack check/push/pop to is_union_sub_x, so that when a union-as-subtype check recurs back to the same pair, it short-circuits instead of re-expanding.

3. Rewrite check_assume to scan the assumption list directly instead of using a push-then-pop probe pattern, eliminating unnecessary ast_dup/ast_free allocations on every check.

Benchmarks on projects using recursive union types (json-ng):

• http_server expr pass: 25.7s → 13.1s (-49%)
• http_server full build: 31.7s → 15.2s (-52%)
• No measurable regression on stdlib or other projects.

[SeanTAllen]

The numbers here are from before I had to change the approach to being more targeted as the first pass caused a crash. I will revisit this.