Qwen3-TTS PR 2: ramped chunk policy and window geometry for streaming consumers

[merceod]

Two small pieces for streaming vocoders (Qwen3-TTS, Qwen3-Omni Code2Wav, Orpheus SNAC, Chatterbox).

1. ScheduledLeftContextChunkPolicy(schedule, chunk, left_context). Left-context chunking whose chunk sizes follow a ramp, so the first audio leaves after a few frames and later chunks grow. LeftContextChunkPolicy needs chunk > left_context and pops 300 frames before the first audio in the Qwen3-TTS deployment (24 s of audio). With schedule=(1, 3, 8, 16), chunk=25, left_context=25 the windows are 1 | 1+3 | 4+8 | 12+16 | 25+25 | ...
2. The consumer gets told the geometry of each popped window instead of re-deriving the producer's schedule. StreamChunk.context_items (leading items an earlier chunk already delivered), num_items and start_offset ride the synthetic streaming edge and land in the consuming request's step_metadata["stream_chunks"][edge], readable from prepare_inputs and cg_key_info (so a consumer can pick its CUDA-graph bucket before its inputs are built). The existing policies report their context too (sliding window window - stride, left context left_context after the first pop, fixed 0), so overlap trimming works for the terminal flush and for any policy.

No behaviour change for existing models. context_items is informational, the new edge fields default to None, and step_metadata["stream_chunks"] is only set for streamed inputs. Nine commits, one per file.

Tests: test/modular/test_stream_chunk_schedule.py (ramped window/context sequence, every item delivered once as new data for 11 stream lengths and two consumer timings, one terminal chunk, context values of the existing policies). Full test/modular passes.