Frame a record into a buffer that lives longer than the record

[bjmeetsfo]

Every append built the payload, then allocated a second buffer to hold the payload plus about ten
bytes of frame, and copied it in. That second allocation is the whole record again, on every write.

The design this is measured against keeps one long-lived operation-log message and a reusable
buffer on its logger rather than minting them per write. This does the same for the frame: the
buffer lives on the log's inner state, encode_record_into clears and refills it, and its capacity
survives.

value	allocations an append	bytes allocated
64 B	34.0 -> 33.0	9,201 -> 9,096
256 B	33.0	10,172 -> 9,873
1024 B	33.0	14,013 -> 12,950
4096 B	33.9	29,390 -> 25,271 (-14%)

The count drops by one; the bytes drop by roughly the size of a record, so the saving grows with
the payload -- which is the right shape, since it removes the cost that scales.

The buffer is taken with mem::take and returned on the success path, so the rest of the inner
state stays usable while it is borrowed. An early return drops it and the next append allocates
once: correct either way, just not free that once. encode_record is unchanged for its other
callers and now allocates a buffer and delegates.

distributed_raft_chunks_large_sequence_add_under_default_entry_limit failed once in the local
full suite and is not from this: it passes 3/3 in isolation on this branch and 3/3 on main, and it
failed the same way on an unrelated branch earlier in the week.