Add SiftQ MiniMax-H3 V2 video provider

[futurestudyspace]

Summary

• Add an independent SiftQ MiniMax-H3 V2 cloud video provider.
• Support text-to-video, first-frame, first-and-last-frame, and reference-to-video modes.
• Support up to 9 reference images plus reference video and WAV/MP3 audio.
• Add a password-masked, session-only API key field that is not stored in .blend files, preferences, metadata, or queue jobs.
• Add queue/Redo propagation, bounded polling, queued-task cancellation, structured redacted errors, validated MP4 downloads, and safe recovery when a create response times out.
• Pass original IMAGE source bytes directly for first-frame, last-frame, and reference-image inputs without baking Blender VSE transforms, crop, resampling, or canvas padding.

Validation

• test_siftq_provider.py: 27 tests passed.
• test_remote_backend.py: 11 checks passed.
• Python compilation checks passed.
• Blender 5.2 extension validation passed.
• Blender 5.2 host verification confirmed that an IMAGE picker resolves to the original source file.
• A live first-frame request produced a valid 5-second MP4.
• Source-image integrity check confirmed that the uploaded data URI decodes byte-for-byte to the original 864×1536 PNG.
• Secret scan found no embedded API keys.

Notes

• MiniMax-H3 is retained only as the model ID required by the V2 protocol. Provider identity, credentials, transport, configuration, implementation, and tests are independent under SiftQ.
• Timeline transforms and crop settings are intentionally not baked into SiftQ IMAGE picker inputs. To use an edited composition, export it as a new image first.