WIP: batched TRT encoder (chunked) — pre-existing, split from the bf16 release

[Cortexelus]

WIP: batched TRT encoder (chunked) — pre-existing

This is pre-existing work-in-progress, split out of the bf16 medium DiT release so that PR stays clean. It was not authored as part of the bf16 change — it happened to live in the same working copy of scripts/sa3_trt_core.py and was separated out hunk-by-hunk.

What it adds

An optional batched path in encode_chunked (scripts/sa3_trt_core.py): when the encoder engine is built with a dynamic batch dimension (optimization-profile max > 1) and the chunk count is in a sweet spot (BATCHED_ENCODER_MIN_CHUNKS ≤ n ≤ BATCHED_ENCODER_MAX_CHUNKS), all chunks are stacked into a single execute_async_v3 call instead of serial per-chunk calls. Adds helpers _runner_max_batch, _runner_supports_batching, _runner_audio_samples_range, _run_batched, and a reworked short-audio single-shot path.

Status — draft

The shipped encoder engine is batch=1, so the batched path is inert until a batched encoder engine is built. The min/max-chunk thresholds and the BF16/FP16-mixed state-drift behavior across super-batches are still being characterized. Opening as a draft to preserve the work and keep it off the bf16 release PR.

Branched off bbd30b5 (same base as the bf16 release).

[Cortexelus]

Pausing the batched-encoder work for now. Closing this PR but keeping the branch batched-encoder-wip with the WIP intact — reopen when we pick it back up.