fix(inference): enable quantized SeaCache and correct Transfer control CFG

[pengcuo]

Quantized inference currently ignores --diffusion-cache, and Transfer does not declare its extra no-control forward to the sampler. Enable the existing SeaCache setting for ModelOpt FP8 and runtime MXFP8/NVFP4, and register the extra branch only while control guidance is active so branch histories and FSDP forward counts stay aligned.

Also sync the control-CFG composition fix: text guidance must scale the existing conditional prediction while the additional control increment remains (control_guidance - 1) * (full - no_control). With text guidance 3 and control guidance 1.5, the old formula subtracts 1.5 times the control delta instead of adding 0.5 times it. Use 1 + (control_guidance - 1) / text_guidance inside the conditional hook and reject zero text guidance when control-CFG is active.

Validation:

• All 43 tests in inference_test.py, transfer_test.py, and diffusion_cache_test.py passed on GB200. Coverage includes distinct cached branch residuals, interval-boundary resets, and control-CFG composition with text guidance. The six new composition/zero-guidance cases fail before the formula fix.
• Ruff lint/format and Pyrefly passed (0 type errors).
• All GitHub checks passed at d91e640, including unit tests, inference smoke tests, and training regressions.
• ModelOpt FP8 cache-off/cache-on generation both produced 121 frames at 1280x720, using 50 UniPC steps, seed 2026, text guidance 3 and control guidance 1.5. Both ran at this PR head with the same checkpoint and inputs, one GPU per run in a four-GPU allocation. Compilation and CUDA graphs were disabled. The cached run recorded 69 full and 81 skipped forwards.
• Synchronized frame comparisons against the edge input show substantially better pose alignment after the control-CFG correction, including removal of the spurious bend at frame 75 in both cache modes. Cache-on versus cache-off PSNR is 35.36 dB; mean adjacent-frame pixel differences are 1.31 and 1.33 respectively (8-bit scale). These are single-clip diagnostics, not a broad quality guarantee.
• Runtime MXFP8/NVFP4 switch behavior has unit coverage; their end-to-end quality has not been revalidated.

Historical reference validation: the saved September 9 no-cache result (245d2b9) used the earlier control-CFG semantics. Its control_guidance=1.5 with text guidance 3 corresponds to control_guidance=2.5 under this formula (new = 1 + text_guidance * (old - 1)). A separate cache-off/cache-on run uses that exact conversion with the original checkpoint, prompts, seed, and control input; the saved control videos decode identically across all 121 frames. Both runs produced 121 frames. The cached result has 32.62 dB mean frame PSNR against the historical no-cache video, and 32.54 dB against the current no-cache result. This explicitly checks the original reported case rather than relying only on agreement between two current outputs.