nemotron_3_diar: write speaker probabilities as safetensors

[LysanderdeJong]

Summary

• return_frame_probabilities writes speaker_probabilities.safetensors instead of raw .f32, with the diarizer's streaming geometry as metadata
• add asr_la0/asr_la3/asr_la6/asr_la13 latency profiles that match Nemotron ASR's chunking at that lookahead
• add engine::io::encode_safetensors() (in-memory, with __metadata__ support); it shares the header builder with write_safetensors_file(), which still streams to disk

Second of three stacked PRs split from #681; #685 (the first) is merged.

Why

The next PR runs NeMo's masked multitalker ASR from this file. That needs the diarizer's geometry: NeMo diarizes in ASR-sized chunks (chunk = lookahead + 1, no right context), and a mask from another geometry is not what the model was trained with. Raw .f32 carries no metadata.

Changes

• The file holds one F32 tensor speaker_probabilities [frames, 8] and string metadata: format_version, source_family, frames, speakers, frame_hop_samples, sample_rate, latency_profile, spkcache_len, fifo_len, chunk_len, chunk_right_context, spkcache_update_period.
• asr_laN sets spkcache 264, FIFO 264, chunk N + 1, right context 0, update period 222, as NeMo's configure_diar_streaming does.
• Docs and spec list the new profiles and file format.

Maintainer notes

• return_frame_probabilities (added in Improve Nemotron 3 diarization NeMo parity #677) now writes speaker_probabilities.safetensors. The tensor data is unchanged. The .f32 output exists only on main since Improve Nemotron 3 diarization NeMo parity #677; no release contains it (v0.8.2 predates Improve Nemotron 3 diarization NeMo parity #677).

Validation

• the tensor is bit-identical to the old .f32 for the same run
• nemotron_3_diar_streaming_test covers the asr_laN geometry and reads the bytes back through load_safetensors_index; the Hugging Face safetensors package reads the file
• existing write_safetensors_file() callers produce identical files

[0xShug0]

@LysanderdeJong Could you elaborate on why you changed the direct-to-disk behavior of write_safetensors_file()? Routing it through the in-memory encoder adds unnecessary full-file buffering and copies.

[LysanderdeJong]

Good catch, that was an unintended regression from a refactor. I only wanted to share the header code. Fixed in 3431e1a: write_safetensors_file() writes directly to disk again, and only the header builder is shared. The in-memory encode_safetensors() is used only for the diarizer artifact, which has to be in memory anyway, and it now builds its buffer in one allocation. Output files are byte-identical.