fix(TextDecoder): don't early-exit the decode loop on predictions sampled during promptTokens prefill

[yangzichao]

Problem

Fixes #372.

With DecodingOptions.promptTokens, the initial prompt becomes [<|startofprev|>, ...prompt..., <|startoftranscript|>, ...] and decodeText force-feeds every one of those tokens through the main loop. The model's sampled prediction at each forced step is discarded (the forced token wins), but it still drives the early-exit checks:

1. If the model's top prediction at any mid-prompt position is <|endoftext|>, sampleResult.completed ends the segment before a single audio token is decoded → empty transcript.
2. isFirstToken points at prefilledIndex (the <|startofprev|> step), so firstTokenLogProbThreshold is evaluated against a forced-prompt placeholder prediction instead of the first real token — this is the spurious Fallback #1 (firstTokenLogProbThreshold) reported in Using promptTokens causes the Transcription to return empty result. #372.

Whether a model predicts EOT at those positions is model-dependent, which is why large-v3-v20240930_turbo and base.en reproduce while base appears fine — but the checks shouldn't be consulting forced-prefill predictions on any model. The reference implementation feeds the prompt without sampling checks and only starts them at sample_begin.

Fix

When promptTokens are present, ignore sampleResult.completed during prefill steps and evaluate the first-token check at the first genuinely sampled position (the last prefill index). The no-prompt path is untouched — checks keep their existing indices, so default behavior is unchanged.

Verification

Standalone harness against the real openai_whisper-large-v3-v20240930_turbo_632MB CoreML model on recordings that reproduced the empty decode:

case	before	after
no prompt	ok	ok (unchanged)
19-token prompt (stripped to content tokens)	empty	ok
single content token	empty (2/3 clips)	ok

With the fix, the prompt also demonstrably conditions the output (a punctuated prompt yields punctuated transcripts), i.e. the channel now works as intended rather than merely not crashing.