Reject trailing Brotli data to prevent infinite drain loops

[illia-v]

Reject trailing data after a complete stream, matching error check of the reference Brotli. This prevents infinite loops while draining the decoder.

Reproducer:

import brotlicffi as brotli  # or import brotli

data = brotli.compress(b"A" * (2**21)) + b"tail"
decoder = brotli.Decompressor()
decoder.process(data, output_buffer_limit=2**20)

while not decoder.can_accept_more_data():
    decoder.process(b"", output_buffer_limit=2**20)

Unpatched brotlicffi loops indefinitely. Patched brotlicffi and reference Brotli raise an error.

[ngoldbaum]

Looks good, this implements a check upstream has had since 2018.

I used an AI model to compare this with upstream and noticed two issues and one test suggestion.

Corner cases worth raising

1. Changelog understates the change. brotlicffi.decompress(stream + junk) and no-limit process(stream + junk) returned the data for the package's entire history and now raise. That's upstream parity and I'd keep it, but HISTORY should say so explicitly rather than only "fixed infinite loops". Note the whole call's output is discarded on the raise, same as upstream.

2. Post-error state isn't reset. Upstream marks the decoder unhealthy. The PR raises before re-saving _unconsumed_data, so when the raise happens on a second loop iteration (any no-limit call whose initial 5 * len buffer is too small, e.g. compress(b"A"*100) + b"tail") a stale tail stays behind:

   1st: Decompression error: trailing data after stream.
   unconsumed: b'tail' can_accept: False finished: True
   2nd process(b'x'): brotli: decoder process called with data when 'can_accept_more_data()' is False
   finish(): b''
   

   Setting self._unconsumed_data = b'' before the raise (or adding a healthy flag) would make the follow-up error consistent. Minor.

3. Tests. Cheap additions that mirror upstream's test_garbage_appended and test_already_finished: one-shot decompress(stream + b'x'), and process(stream) then process(b'x').

[illia-v]

@ngoldbaum thanks for the review! I updated the changelog, fixed the stale buffer, and added the regression tests