Honour CPU offload on H3, fix the training status, unblock large HF downloads

[imprsnst]

Three fixes from field reports, plus the version bump.

CPU offload was silently dead on H3. resolve_offload tested the quantisation before the user's explicit choice, so On returned False for any 4-bit base. H3 is always 4-bit, so the one control that helps its users did nothing. Reordered: on and off win outright, auto keeps the old conservative behaviour. H3 breaks the "a quantized base already fits" assumption, because its base is 11.7GB and it is the clip activations that overflow the card.

Progress was only emitted after a step finished, so the UI showed "loading model (nf4)" for the whole of step one. A slow first step read as a hung loader and sent a reporter and me chasing bitsandbytes. Now announced before the loop, with a test pinning the order.
Possible fix to #32

Large HF downloads failed. hf_xet became a hard dependency of huggingface_hub at 0.32, and our floor was 0.23, so an older resolve could not fetch the 66GB H3 transformer at all. Floor raised.

Also carries the --no-sources launcher fix: --no-sources-package only exists in uv 0.10+, so the installer hard-errored for anyone older.