Factor: collapse triplicated alloc recovery into one arm

[cds-amal]

This is a follow-up to #131. That PR fixed a real panic (non-builtin-deref Static/VTable allocations), but the fix landed as three nearly identical match arms: Static, VTable, and Function, each repeating the same "try get_prov_ty, fall back to opaque placeholder" logic. The only meaningful difference between them was a single predicate ("is this type usable directly?") and the debug log string. This PR collapses them back into one arm, makes the predicate explicit, and documents a previously unexplained jq filter.

What changed

src/printer/mir_visitor.rs

The three arms now share a single code path. The predicate that actually differs between them is captured in a needs_recovery variable:

• Static/VTable: builtin_deref(true).is_none() (not a reference, raw pointer, or Box)
• Function: !kind.is_fn_ptr() (outer type isn't already a function pointer)

Worth noting: these two predicates are not equivalent (a function pointer passes builtin_deref but fails is_fn_ptr), which is why a naive "just combine the arms" without the inner match would have changed behavior for Function allocs. The inner match on global_alloc makes this asymmetry visible rather than hiding it across separate arms.

The rest of the logic (try get_prov_ty, fall back to opaque placeholder) is written once. Also dropped an unnecessary .clone() on the builtin_deref call; it takes &self.

tests/integration/normalise-filter.jq

The def_id filter had a terse comment ("unrelated to this regression") that didn't explain why it exists or why it's safe to strip globally. Replaced it with a proper explanation: def_id values are interned compiler indices (same class as alloc_id and adt_def) that are consistent within a single rustc invocation but non-deterministic across runs. Downstream consumers need them as cross-reference keys to join AggregateKind::Adt in MIR bodies with type metadata entries, so they can't be dropped from the output itself; we only strip them here for golden-file comparison. (See #125 for the full picture on interned-index non-determinism.)

Test plan

• cargo build compiles cleanly
• make integration-test passes (no behavioral change; this is a pure refactor)

[dkcumming]

Nice! Tested downstream with mir-semantics (682df6b2dfec4ec0516e246f22b70dff09cf57a6) test suite which is passing