feat(backends): add LongCat video and avatar generation

[localai-bot]

Description

Adds a dedicated longcat-video Python gRPC backend for LocalAI, supporting LongCat-Video text-to-video, image-to-video, and LongCat-Video-Avatar 1.5 audio-driven avatar generation.

The change extends /video with safely staged audio and video-generation parameters, wires distributed audio staging, supports CUDA 12/13 x86_64 and CUDA 13 ARM64 backend images, and exposes audio/reference-image controls in the React Studio UI.

Capability design

• Adds declarative known_input_modalities and known_output_modalities model-config fields.
• Capability discovery merges declared modalities with existing inference in canonical order.
• LongCat gallery and importer recipes declare their actual I/O contract; generic model-capability code no longer branches on the LongCat backend or checkpoint name.
• The model configuration editor and API-discovery documentation expose the same declarations.

Model gallery

• longcat-video: text + image input, video output.
• longcat-video-avatar-1.5: text + image + audio input, video output.
• Hugging Face imports emit the same self-describing recipe metadata.

Documentation

Adds a dedicated LongCat Video guide covering gallery installation, DGX Spark and NVIDIA ARM64 guidance, BF16/SDPA versus INT8 tradeoffs, Studio usage, text-to-video, image-to-video, avatar API examples, load options, request parameters, troubleshooting, and current limits. The general backend, video-generation, model-configuration, API-discovery, feature index, and whats-new pages link to it.

Backend notes

• Pins the upstream LongCat-Video source and applies a minimal SDPA compatibility patch verified against that exact revision.
• Avatar mode supports multi-segment generation, BF16, optional INT8 quantization, and distillation settings.
• Input staging is bounded to 128 MiB with redirect, HTTP-status, path, and cleanup handling.
• Fixes the patch context that caused the first PR backend-image jobs to fail; replacement CUDA 12/13 amd64 and CUDA 13 arm64 jobs are running.

Testing

• Full pre-commit make lint and make test-coverage-check; composite Go coverage 52.1% versus the 48.5% baseline.
• Pre-commit make test-ui-coverage-check from the initial implementation (215 Playwright tests; UI line coverage 47.95%).
• python3 backend/python/longcat-video/test.py (10 tests, 1 skipped).
• Exact pinned-upstream patch apply check plus Python byte-compilation of both patched modules.
• Targeted config-capability, importer, video-endpoint, and distributed-staging tests.
• React production build, focused Playwright tests, Ruff, shellcheck, generated protobuf, Swagger, YAML/JSON, and requirements validation.

Full checkpoint inference and a complete backend container build were not run locally: the model set requires roughly 83 GiB of downloads and NVIDIA CUDA/DGX-class hardware. CI is responsible for the backend image builds.

Signed commits

• Yes, I signed my commits.

The human submitter should review and add the required DCO sign-off before merge.