Batched generation: generate_batch() with ragged-batch S3Gen

[TroyHernandez]

Adds generate_batch(model, texts, voice, ...): T3 token generation runs per text (autoregressive generation doesn't batch — lengths and EOS differ), then ALL utterances synthesize in one batched S3Gen pass — a single CFM solve and a single vocoder call over the padded batch, trimmed per row. This is the upstream 0.1.7 speech_token_lens batching, not pipelining.

Plumbing changes:

• s3gen$inference() accepts (B, T) tokens + speech_token_lens for ragged batches
• solve_euler CFG generalized from hardcoded batch-2 to 2B (traced falls back to non-traced for B > 1)
• flow builds per-row mel masks and expands single-voice conditioning across rows
• generate()'s T3 stage extracted to .t3_text_to_tokens, shared by both entry points

Three padded-batch leaks found and fixed (each invisible at B = 1):

1. ref_dict's (1, 1)-shaped prompt_token_len broadcast a (B) length vector to (1, B), collapsing the batch in make_pad_mask
2. the CFM estimator's transformer blocks ran unmasked — padded tails bled into valid frames through bidirectional attention; a key-padding mask is now built when padding exists
3. the conformer pre-lookahead conv reads rightward and pulled nonzero embedded-padding into the last valid frames; the tail is zeroed after embed

Validation (GPU, identical token sequences): batch row vs single run — encoder ≤ 2e-4, mel through the full 10-step CFM 0.003–0.005 (the single-run FP envelope; the Python-parity bar is 0.03). End-to-end batch of 3 texts: all EOS, correct durations, samples in ~/Sync.

CPU regression test for the mask/broadcast math (test_batch_masks.R); full suite passes. Version 0.1.0.8 + NEWS as separate bump commit.