feat(minimax-remover): add NVFP4 kernelized video inpainting pipeline

[chenping9999]

Performance highlights (RTX 5060 Ti, SM120, CUDA 13):

• End-to-end: 2.6x speedup (30.7s -> 11.9s for 123 frames, 3 segments)
• RTF improvement: 6.0 -> 2.3 (processing-time / clip-duration)
• Per-layer GEMM: 1.14-1.30x speedup vs fp16 matmul
• FP4 GEMM with quantize overhead: 4-9x faster than fp16 on large FFN projections
• Precision: PSNR 52.0 dB (mean) / 45.2 dB (worst frame) vs fp16

Optimizations:

• NVFP4 W4A4 quantization with dynamic per-call activation (no offline calibration)
• Fused LayerNorm + adaLN + gate-residual in single fp32-stat Triton kernel
• Fused FFN-up GEMM + bias + GELU -> FP4 output (skip re-quantization for FFN-down)
• QKV quantize-once optimization (reuse quantized norm output for Q/K/V projections)
• Manual graph-capturable denoise loop with CUDA Graph support
• BF16 transformer (NVFP4-native, eliminates fp16<->bf16 casts)
• SageAttention / FA2 backend integration
• Shared attention dispatch for installed processors and the manual fused block, so FLASHRT_ATTN_MODE=fa2 uses the dependency-light FA2 path without importing sageattention

Validation:

• PYTHONPATH=. PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 pytest -q tests/test_minimax_remover_smoke.py: 8 passed, 2 skipped
• python -m compileall -q flash_rt/models/minimax_remover tests/test_minimax_remover_smoke.py examples/minimax_remover_quickstart.py: passed
• git diff --check: passed
• Import check for flash_rt, flash_rt.models.minimax_remover, _kernels, and pipeline: passed
• FLASHRT_ATTN_MODE=fa2 dispatch check with a stub flash_rt_fa2: passed without touching SageAttention

[LiangSu8899]

Thanks for the contribution. The direction is useful: adding MiniMax-Remover as an isolated model package, reusing existing generic NVFP4 kernels instead of adding model-specific CUDA operators, is the right long-term shape.

Review result: Request changes / blocker before merge.
Risk class: R3/R4, because this adds a new model runtime path with optional dependencies, NVFP4 required-symbol checks, Triton kernels, and performance claims.

I do not see evidence that this affects the default FlashRT core path directly: the PR adds new files only and does not modify CMake, bindings, or existing model routing. That part is good. The blockers are mostly packaging/import/docs/test contract issues.

Before merge, please fix these items:

• Keep optional dependency boundaries intact. import flash_rt.models.minimax_remover should succeed without diffusers, sageattention, or the MiniMax-Remover reference package installed. Right now package import reaches flash_rt/models/minimax_remover/_manual_denoise.py, which imports diffusers at module import time and fails in a default environment.
• Move optional imports to runtime boundaries or fail-fast constructors. Dependencies such as diffusers, einops, triton, and sageattention should either be in a documented optional extra or imported lazily with a clear error message when the MiniMax-Remover path is actually constructed/used.
• Add a package extra or clear install command for this model path. For example, a dedicated extra like minimax-remover = [...] or an equivalent documented command. The current pyproject.toml does not advertise the dependencies needed by this new path.
• Fix the build documentation. ENABLE_CUTLASS_SM120_NVFP4_W4A16 is an internal compile definition, not the public CMake entry point users should pass. Please document the real build route, e.g. Blackwell GPU_ARCH=120 / 121 and the existing NVFP4 build conditions.
• Make tests/test_minimax_remover_smoke.py pass in a default environment without diffusers installed. The test comments say import should succeed and _load_kernels() should fail clearly only when the kernel surface is missing, but currently the tests fail at import time.
• Clarify the default attention dependency. The docs say the default is FLASHRT_ATTN_MODE=sage_fp8, which requires sageattention. Either make that dependency part of the model extra, make fa2 the dependency-light default, or fail fast with an explicit install hint.

What I verified locally:

• git diff --name-status origin/main...origin/pr123-head: only new MiniMax-Remover docs/model/test files are added.
• git diff --check origin/main...origin/pr123-head: passed.
• python -m compileall flash_rt/models/minimax_remover tests/test_minimax_remover_smoke.py: passed.
• No private paths, debug breakpoints, or AI-generated traces found in the changed files.
• In an environment without diffusers, PYTHONPATH=. python -c "import flash_rt.models.minimax_remover" fails with ModuleNotFoundError: No module named 'diffusers'.
• In the same environment, PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 PYTHONPATH=. pytest -q tests/test_minimax_remover_smoke.py reports 6 failed, 2 skipped for the same import-boundary issue.

For the maintenance standard we are applying, please refer to the public review checklist:

https://github.com/flashrt-project/FlashRT/blob/main/docs/pr_review_checklist.md

The relevant sections are mainly import boundaries / optional dependencies, new model acceptance, docs matching actual build flags, and default-environment tests. Once these are fixed, the PR will be much easier to review on the actual model/runtime behavior.