TorchScript greedy decode step + CUDA GC tuning

[TroyHernandez]

Applies the GC and JIT speedups from the recent chatterbox work to whisper.

JIT greedy decode (R/decode_jit.R)

Each generated token's full decoder forward (per layer: self-attention, cross-attention, FFN; plus the final LayerNorm) now runs as one jit_compile'd TorchScript call. R keeps the eager prefill, the sampling/timestamp logic, and the loop shell.

Same motivation as chatterbox's t3_inference_jit: even optimally-written eager R hits a per-op R->lantern dispatch floor (hundreds of ops/token). Collapsing the per-token forward into one libtorch call removes it without compiled code or linking against torch's private libraries.

Whisper's step differs from a Llama step in three ways, all handled: LayerNorm with bias (hand-rolled in float for fp16 parity), biases on every projection, and a cross-attention block whose encoder K/V are cached once and passed in as stacked tensors.

• On by default via the new jit arg to transcribe() / whisper_pipeline().
• Gated to CUDA greedy non-word-timestamp runs. CPU, beam search, and word timestamps stay on the eager decoder, so R CMD check (CPU) is unchanged.

GC tuning (whisper_tune_gc())

Opt-in helper that raises torch's CUDA allocator GC floor (torch.cuda_allocator_reserved_rate) to the model's footprint as a fraction of VRAM and lifts torch.threshold_call_gc off its default, so GC stops firing on nearly every allocation during inference. Call before load_whisper_model(). No-op off CUDA; only sets options that aren't already set.

Verification

• Token-for-token identical to eager greedy, with and without timestamp rules (logprob diff 0.006, pure fp16 rounding). New test_decode_jit.R (CUDA + at_home gated).
• ~2.5x faster end-to-end on medium for the jfk clip; more on longer transcripts where per-token dispatch dominates. Text identical.
• Full suite: 199 results, all OK.

Version 0.3.0 -> 0.3.0.1.