Make gradient checkpointing and offloading per-component

[dxqb]

• fixes [Bug]: Activations offloading depends on layer offload fraction #1136 and [Feat]: Separate offload settings for text encoder #980

• includes fix: Flux2 UI — sequence length and caption dropout labels no longer overlap #1460 because it was in the way

• required for Microsoft Lens, because it has a text encoder that cannot be (easily) offloaded

• some general testing

• testing of SD VAE training because this PR touches it and it's rarely ever tested

[dxqb]

Claude:

• activation_offloading should default to False — it was always default-True but only took effect with layer offloading; now it works standalone, so a fresh fine-tune offloads activations out of the box.
• "Layer Offload Fraction" is shown for CLIP encoders but is a no-op — the CLIP setup discards its conductor, so the value never drives anything; don't render the field for CLIP (or wire it up).

[dxqb]

Claude: Found while testing the preview branch — the caching_threads > 1 + layer-offloading guard in create_data_loader (introduced here, with a TODO: narrow this to the cached components only) is now overly broad:

if config.caching_threads > 1 and any(part.offload_fraction > 0 for part in config.model_part_configs()):
    raise RuntimeError('layer offloading can not be activated if "caching_threads" > 1')

This rejects the config if any part has layer offloading enabled — including transformer/unet/prior/unconditional_transformer, none of which run inside the caching dataloader's worker threads (only the components that actually produce a cache do: text encoder(s) for text caching, VAE for image/latent caching).

In practice, after the per-component split:

• Only text_encoder / text_encoder_2 / text_encoder_3 / text_encoder_4 (depending on model_type.model_parts()) expose an "Offload" UI control (__create_offloading_widgets in BaseTrainingTabView.py) — vae does not (__create_vae_frame has no offloading widgets), so vae.offload_fraction is always 0 in practice.
• So the check should really be: does any text encoder part that's actually used (and being cached) have offload_fraction > 0?

This means the current check blocks perfectly valid configs — e.g. layer offloading the transformer with caching_threads > 1 — even though that combination is fine, since the transformer's conductor never runs in a caching worker thread.

Suggested narrowing: only check text-encoder parts, e.g.

if config.caching_threads > 1 and any(
    getattr(config, name).offload_fraction > 0
    for name in config.model_type.model_parts()
    if name.startswith("text_encoder")
):
    raise RuntimeError('layer offloading can not be activated for a text encoder if "caching_threads" > 1')