ModelDumper output buffer persists across calls

[yarikoptic]

Reusing a single ModelDumper instance across two `dumps()calls returns the **concatenation** of both inputs on the second call. This makes the model API unusable for any program that processes more than one document, and likely explains whytests/test_roundtrip.py`` is
fully commented out upstream.

Reproduce, just run

from json5.dumper import ModelDumper, dumps
from json5.loader import ModelLoader, loads


dumper = ModelDumper()  # SINGLE instance, reused below

a = loads('{"a": 1}', loader=ModelLoader())
b = loads('{"b": 2}', loader=ModelLoader())

out_a = dumps(a, dumper=dumper)
out_b = dumps(b, dumper=dumper)

print(f"first  dumps -> {out_a!r}")
print(f"second dumps -> {out_b!r}")

and get

    first  dumps -> '{"a": 1}'
    second dumps -> '{"a": 1}{"b": 2}'

I thought that may be it is just by design to be cumulative across records... but unlikely since

a = loads('{"a": 1, "c": 4}', loader=ModelLoader())
b = loads('{"a": 2, "b": 2}', loader=ModelLoader())

out_a = dumps(a, dumper=dumper)
out_b = dumps(b, dumper=dumper)

print(f"second dumps -> {out_b!r}")

leading to not really sensible

second dumps -> '{"a": 1, "c": 4}{"a": 2, "b": 2}'

It was discovered while claude code round-tripping 2,151 sidecars from the https://github.com/bids-standard/bids-examples/ which are quite a wild west of formattings of jsons there. Good news is that after working around this issue, it reported that all round trips succeeded.