Allocate block-exceeding tokens directly in the old generation

[Robertorosmaninho]

While running large MegaETH workloads under the LLVM backend we hit a crash in kore_collect / arena::evacuate while migrating a block with a corrupted header. The culprit is a String token larger than a block: init_with_len correctly flags it NOT_YOUNG, but kore_alloc_token was still allocating it in the young semispace. The collector never migrates a NOT_YOUNG object out of the young from-space, which is reclaimed on every collection, so the token dangled — a later allocation overwrote its bytes and the next collection walked a garbage header. It reproduces independently of the hyperblock size.

The fix routes tokens larger than BLOCK_SIZE to oldspace (matching what hook_BUFFER_concat_raw already does for large buffers) and sets AGE_MASK on the NOT_YOUNG header so is_in_old_gen_hdr() recognises the block as a live old-generation object. The oversized branch is kept in a cold, out-of-line helper (__builtin_expect) so the common small-token path of the always_inline kore_alloc_token isn't bloated at every call site.

I also added test-gc-large-token: it builds a single > 1 MiB token, keeps it live across forced collections that reuse the abandoned young region, then reads it back. The test crashes on the buggy version (the collector walks the dangling token) and passes with the fix; the rest of the GC and unit suites are unaffected.

This is safe for the hot path: small-token allocation is unchanged apart from one predicted-not-taken branch, and the oversized path is rare. Instruction counts on an allocation-heavy benchmark are flat versus before, and the full lit suite stays green.

[gtrepta]

LGTM