Fix end-of-turn token leaking into streamed chat output

[PratikDhanave]

What

_print_stream in gemma/gm/text/_chat_sampler.py (used by ChatSampler on the streaming path) is meant to suppress the end-of-turn marker from the streamed/returned text — the comment says "Last token is not printed." But the guard is placed after the work it should gate:

for state in out:
    print_(stream, state.text)          # printed before the check
    text_tokens.append(state.text)      # appended before the check
    if state.text == '<end_of_turn>' or state.text == '<turn|>':
        continue                        # no-op: last statement in the loop body
out = dataclasses.replace(state, text=''.join(text_tokens))

Because print_ and append run before the if, and continue is the final statement in the loop body (so it does nothing), the end-of-turn marker (<end_of_turn> / <turn|>) is:

1. streamed to the user, and
2. concatenated into the returned SamplerOutput.text.

Any streamed chat turn ending normally therefore leaks the literal end-of-turn token into the user-visible output.

Change

Move the guard ahead of the print/append so the marker is skipped, matching the documented intent. No other behavior change.

for state in out:
    if state.text == '<end_of_turn>' or state.text == '<turn|>':
        # Last token is not printed.
        continue
    print_(stream, state.text)
    text_tokens.append(state.text)