Tell the key buffers their size, not an offset past their end

jolt.ffi/write took (ptr type offset value) and now takes (ptr type
value offset), with the offset trailing and optional — the babashka.ffi
order, changed upstream in 4feca766. Jolt moved its own callers in that
commit; these two did not move, and nothing said so, because the arity
is still legal and both arguments are still integers.

So `(ffi/write plen :size_t 0 32)` stopped meaning "this buffer is 32
bytes" and started meaning "store 0 at offset 32" — an eight-byte write
thirty-two bytes past the end of an eight-byte block, immediately
before EVP_PKEY_get_raw_public_key. slen is the same store sixty-four
bytes out. What that cost was a heap glibc noticed later and elsewhere:
`malloc(): unaligned tcache chunk detected`, aborting inside OpenSSL's
own allocator on the TLS dial, with a backtrace that names ecx_export
and nothing of frq's.

The runtime bump two commits ago did not introduce this; it uncovered
it. read still takes its offset third, and glimmer, glimmer-vidya and
glimmer-tui call neither, so these were the only two sites.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>