Respect caller timestamp options in batched transcribe

# Summary This PR restores caller-provided `condition_on_previous_text` and `max_initial_timestamp` in `BatchedInferencePipeline.transcribe()`. Related issue: #1427 Previously, the batched path hardcoded these values instead of using the caller request, making it behaviorally non-equivalent to `WhisperModel.transcribe()` for timestamped decoding. ## What this PR changes - pass through `condition_on_previous_text` in batched transcription options - pass through `max_initial_timestamp` in batched transcription options - add regression tests that verify both values are preserved in batched transcription options ## What this PR does not try to solve This PR is intentionally narrow. It does **not** attempt to fix all direct/batched quality gaps such as: - prompt/history update design - fallback behavior differences - VAD chunk packing strategy ## Validation - added lightweight regression tests for option propagation - verified the new tests pass locally with `pytest` - confirmed on a public FLEURS Japanese fixture that upstream batched transcription reports the overridden values, while the patched branch preserves the caller values