A pure main forces deep_thunk_chain's 262144-link Nat thunk chain in the source interpreter, which its IO main left to the JS lane

[HamedShams]

Closes #1316.

tests/eval/deep_thunk_chain.bend pins the interpreter collapsing a lazy accumulator deeper than the host stack, but its main is IO, and bend file.bend runs an IO main through Comp.io_run, i.e. the JS emitter. So the chain that comment describes only ever gets forced by the strict JS lane; the interpreter never sees it. Only a pure main goes through term_snf.

This adds tests/eval/deep_thunk_chain_pure.bend: same Nat.pow2, same go, same head demand through a match, but main answers a Bool instead of printing. The interpreter now has to collapse the 262144-link Nat.add(acc, 1n) chain itself before the match can pick an arm. I kept the match rather than returning the Nat so the test pins thunk forcing alone, not the printer's walk over a 262144-deep Succ chain, which is a separate concern.

Ran it the way the gate does, from the repo root:

• --check-only: ALL PROOFS CHECK
• aggregator --checkup with BUN_JSC_maxPerThreadStackUsage=33554432: prints True{} under the file header, 1.1 s
• -o t.js + bun, and -o t + ./t: both True{}, well under the 5 s alarm
• bun gates/repo.ts: PASS

It also passes with a 512 KB JSC stack, so the forcing really is off the host stack. I couldn't rerun it from before the LAZ-frame law landed: the tree at that point predates law, so the parser rejects the file.