Review mflux main through #828 in the drift guard, with float16-compute parity

[IonDen]

The test job that runs against mflux's main branch went red after mflux-community/mflux#820, which adds an opt-in float16 compute mode for attention and feed-forward on FLUX.2, Z-Image and Qwen-Image 2.1. This PR reviews that change and moves the drift guard's reviewed commit forward. It changes tests only, so there is no release.

#820 changed two functions TeaCache wraps, Flux2Klein.generate_image and ZImage.generate_image, and in both only the image-metadata call: it now passes the compute setting as generation_parameters (in Z-Image, next to the existing float32 entry). TeaCache passes that return value through. The float16 setting itself lives on mflux's attention and feed-forward modules, which TeaCache calls rather than copies, so a TeaCache run picks it up as plain mflux does. No code TeaCache copies reads it.

To confirm that, the weight-free parity test now builds small real Z-Image and FLUX.2 transformers with the float16 option on and checks that TeaCache's gated forward at threshold 0 is bit-identical to the transformer's own call, with and without CFG. The cases skip on mflux releases that do not have the option. On FLUX.2 the tiny transformer's block weights are scaled up so float16 compute actually changes the output; at the default initialisation the two were identical and the check would have proved nothing, and a guard test keeps that from regressing.

The reviewed commit moves from 04fc101 to e50215c, which also contains #824, #813 and #828. Those three are off TeaCache's paths: #824 changes how the initializers load a baked LoRA, #813 adjusts test tolerances and #828 a CLI warning. The two generate_image fingerprints are the only ones that changed. On a git install of mflux, the parity cases now fail instead of skipping if the bfloat16 stream or the float16 option disappears, so an upstream rename cannot silently turn the check off. The fast tests pass on mflux 0.21.0 and on mflux main at e50215c.