Batched encoder: run N clips through one fused ggml graph

[mudler]

Adds batched-encoder support: N audio clips run through ONE fused ggml encoder graph (subsampling, conformer stack, output) at batch size B, instead of one graph per clip. The single-clip (B=1) path is kept numerically faithful to before. Decode stays per-item over the batched encoder output (batched decode is intentionally a separate future milestone).

What landed:

• MelBatch + Encoder::forward_batch (one fused batched graph).
• Batched builders, each with B on the trailing axis and a per-item valid_len vector driving the masks; every scalar build_graph is now a thin B=1 adapter (DRY, no duplicated bodies): Subsampling::build_graph_batched, RelPosAttention::build_graph_batched (incl. the 4D rel-shift), ConformerLayer::build_graph_batched, and a B-aware build_conv_module.
• Model::transcribe_pcm_batch (+ extracted decode_enc_out) and C-API parakeet_capi_transcribe_pcm_batch.
• bench-batch CLI subcommand + BENCHMARK.md notes.

Correctness

Built test-first against the smallest validated model (parakeet-tdt_ctc-110m, f32 so the tight 1e-2 parity tolerances hold). Per-item batched-vs-standalone equivalence + padding-invariance tests at every level: test_subsampling_batch, test_relpos_attention_batch, test_conformer_batch, test_encoder_batch, test_capi_batch.

• Attention and encoder equivalence are bit-exact (max|d| = 0) per item, including ragged batches; the full encoder is within ~2e-6 over 17 layers. Padding invariance is explicitly asserted (a shorter clip does not perturb a longer neighbor).
• B=1 parity vs the NeMo baseline is unchanged across subsampling/relpos/conformer/encoder.

Two real bugs were caught by these gates and fixed:

1. Subsampling padding leakage: the non-causal path only masked the flattened output; in a batch the last valid frame read the bias-contaminated zero-pad region. Fixed with per-stage per-item trailing-pad input masking, matching NeMo's MaskedConvSequential.
2. Conformer depthwise conv: ggml's 1D im2col requires ne[3]==1, so it aborted for B>1. Fixed by running the depthwise per item and concatenating; B=1 is byte-identical.

Performance note (important)

On a CPU backend, batching does not improve throughput and slightly reduces it (B=1 is fastest): the GEMMs already saturate the thread pool at B=1, and padding every clip to the batch's longest clip just adds wasted work. The throughput win is a GPU-backend effect. Documented in BENCHMARK.md and surfaced by the new bench-batch subcommand. On CPU the value of this PR is the correct, fused batched encoder graph; the latency/throughput payoff lands on GPU.

Scope / limitations

• Encoder only; decode is per-item (batched decode is future work).
• Causal/streaming subsampling at B>1 is hard-guarded (GGML_ASSERT); only the offline non-causal path is batched.
• No length bucketing (mixed-duration batches waste pad compute, same as NeMo's own transcribe).

Test plan

• ctest -L model green: 100%, 0 failures (38 tests, with a converted 110m f32 model + NeMo baseline).
• All 5 batch tests pass; B=1 parity unchanged vs baseline.