fix: -t with -m sums timestamps instead of taking the newest

[VXNCXNX]

-t (group by file type) combined with -m (show filetimes) adds the timestamps
together, so every group shows a date a century in the future.

Three .log files with mtimes 2020-01-01, 2021-06-15 and 2022-03-10:

$ dust -P -c -w 60 -t -m m demo
before:
2123-08-24T12:00:00   ┌──.log
2123-08-24T12:00:00 ┌─┴(total)

after:
2022-03-10T12:00:00   ┌──.log
2022-03-10T12:00:00 ┌─┴(total)

The same tree without -t already prints 2020, 2021 and 2022 correctly, which is
the reference the grouped view should agree with.

Cause

When by_filetime is set, node.size holds a timestamp rather than a size, and
get_all_file_types already knows that in two places:

let actual_size = if by_filetime.is_some() {
    ext_nodes_iter.map(|node| node.size).max().unwrap_or(0)
} else {
    ext_nodes_iter.map(|node| node.size).sum()
};

The per-extension accumulator in build_by_all_file_types was not given the same
treatment, so it keeps doing *cumulative_size += node.size and sums epochs.

The fix

Pass by_filetime down and take the max instead of the sum in that one branch,
matching what the surrounding code already does. Size mode is untouched.

Verification

test_show_files_by_type_with_filetime builds the three files above and asserts
the rendered date.

Reverting only the branch back to += fails it, and the failure output is the
symptom itself:

test test_show_files_by_type_with_filetime ... FAILED
2124-06-30T12:00:00   ┌──.log
2124-06-30T12:00:00 ┌─┴(total)

I kept a test here rather than going test-free as on #609: this one needs real
files on disk, since build_by_all_file_types only counts nodes where
node.name.is_file(), so it is not the kind of one-line config change that
reads correct on sight. The timestamps are set at midday UTC so the assertion
holds regardless of the runner's timezone.

cargo test is 31 + 8 + 32 + 4 passed, 0 failed. cargo fmt --check clean.
cargo clippy reports the same two pre-existing "empty line after doc comment"
warnings before and after, in files I did not touch.

Disclosure: written with AI assistance (Claude Code). I built binaries from master and from the patched tree and produced the before and after above by running them, and ran the mutation check myself.