Add parakeet_capi_transcribe_pcm_logits: expose CTC log-probs

[aanchan]

Summary

The C-API currently only ever returns decoded text/timestamps/JSON, even
though CTCDecoder::forward already computes a log-softmaxed [T, vocab+1]
matrix internally (src/ctc_decoder.cpp, via ggml_soft_max -> ggml_log).
Anyone doing external LM fusion on top of this library's CTC output (e.g.
pyctcdecode + a KenLM n-gram LM + hotword boosting) currently has no way to
get that matrix — only the greedy/beam text this library's own decode
produces.

This PR adds one new C-API entry point, parakeet_capi_transcribe_pcm_logits,
that runs mel + encoder + the CTC head on a PCM buffer and returns the
log-prob matrix (row-major [T, vocab+1], already log-softmaxed) instead of
decoded text, plus parakeet_capi_free_logits to release it. ABI bumped to
v6; all existing entry points are unchanged.

What's included

• Model::transcribe_pcm_ctc_logits / Model::transcribe_16k_ctc_logits in
  src/model.hpp/src/model.cpp: mirrors transcribe_pcm/transcribe_16k
  through mel + encoder (including the existing long-audio tiling path — no
  new truncation limit), then runs only the CTC head, skipping greedy/beam
  decode. Always forces the CTC head regardless of the model's preferred
  decoder, and throws std::runtime_error (via the existing
  ctc_head_tensor lookup) if the model has no CTC head at all — e.g. a
  pure TDT/RNNT-only streaming model.
• parakeet_capi_transcribe_pcm_logits / parakeet_capi_free_logits in
  include/parakeet_capi.h / src/parakeet_capi.cpp, following the same
  validate -> try/catch -> last_error conventions as every other capi_*
  function (no C++ exception ever crosses the C boundary; a NULL ctx or
  any NULL out-param returns nonzero without writing through any pointer,
  otherwise all outputs are zeroed up front so every failure path leaves
  defined state).
• tests/test_capi_ctc_logits.cpp (registered in tests/CMakeLists.txt,
  LABELS "model"), two independent blocks:
  1. Self-consistency on a real standalone-CTC checkpoint
     (PARAKEET_TEST_GGUF_CTC), through the actual C-API (not just the
     underlying C++ method): argmax-greedy decode over the exposed matrix
     (via the same ctc_greedy + detokenize path decode_enc_out uses
     internally) reproduces transcribe_pcm(..., kCTC)'s own greedy
     transcript byte-for-byte. Verified against the real
     parakeet-ctc-1.1b q8_0 checkpoint on tests/fixtures/speech.wav.
     Only blank_id/tokenizer_pieces are cached from the reference
     pk::Model before it's released, so a second full model isn't kept
     resident alongside the C-API's own load.
  2. Error-path coverage at the C-API boundary
     (PARAKEET_TEST_GGUF_NO_CTC): a model with no CTC head (verified
     against the 120M realtime-EOU streaming model) fails cleanly —
     nonzero return, *out_logits left NULL, last_error set — never a
     crash.

Motivation / context

We're building a second parakeet.cpp-backed compute backend for our own
CTC decode stack (pyctcdecode + KenLM + hotwords) that already runs against
another CTC engine, and this is the one piece of surface area this library
was missing to slot into it. Filing upstream since any consumer doing
external LM fusion on this library's CTC output hits the same gap.

This has already been through two rounds of review on our own fork
(happy-prime-inc#1) with fixes applied for validation
ordering, output-zeroing on error paths, test peak-RAM, and a couple of
comment-accuracy nits — landing here in what should be a reviewable state,
not a first draft.

Happy to adjust the entry point's shape (out-param ordering, naming, etc.)
if you'd prefer something different from what's here.

[localai-org-maint-bot]

@mudler This is good to merge from my review. The new C entry point preserves the existing mel/encoder/prompt-conditioning path, returns the CTCDecoder post-log-softmax [T, vocab+1] layout with a symmetric free function, and keeps failures inside the C boundary with cleared output ownership. The focused test covers C-API allocation/free self-consistency and the no-CTC error path. Source inspection and diff hygiene pass; I could not execute the local CPU build because this host currently has no cmake binary.