fix(mem_forest): return writer errors from serialize instead of panicking

[USCMig]

MemForest::serialize panics on a failing writer

Rebased onto main. The deserialize fixes this PR originally carried landed
independently in 69484de ("apply fixes for bugs found while fuzzing"), so that
commit is dropped. What remains is the one fix main doesn't have yet.

serialize returns io::Result<()> and ?s its two length prefixes, but then
unwrap()s the result of writing each root:

root.write_one(&mut writer).unwrap();

A writer that fails partway (a full disk, a closed pipe, a socket that went
away) panics instead of returning the Err the signature promises. The fix is
?.

Testing

test_serialize_reports_writer_errors serializes a forest into a writer with
room for the two length prefixes but not for the roots. It expects
ErrorKind::WriteZero, then checks that the same forest still serializes into
an unbounded writer.

Reverting the fix makes exactly this test fail, and it panics at
src/mem_forest/mod.rs:271.

cargo test --all-features passes (105 tests). cargo clippy --all-targets --all-features is clean, and so is cargo +nightly fmt --check. There is no
public API change, and no behaviour change for any writer that succeeds.

A note on 69484de, not a request

main bounds _read_one at depth > MAX_FOREST_ROWS + 1. A root has at most
MAX_FOREST_ROWS rows, so its leaves are at most that many levels down, and
> MAX_FOREST_ROWS is the tight bound. The + 1 is harmless, since the
recursion is bounded either way. I'm mentioning it only because the tighter
form was in this PR before the rebase.

[Davidson-Souza]

Needs rebase.

[USCMig]

new commit to cleanly merge

[Davidson-Souza]

CI is red.

[USCMig]

Fixed CI issue and it green, ready to merge. Thanks!