lora: make lora_strength scale DoRA/BoRA magnitudes, not just the direction

[Cortexelus]

Problem

lora_strength scales the low-rank direction update, but a DoRA/BoRA magnitude vector replaces the base row/column norm outright. So as strength → 0 the direction converges to W0/‖W0‖ while the magnitude stays fully learned, leaving (mag/‖W0‖)·W0 — whereas strength == 0 short-circuits to exactly W0.

That is a discontinuity at zero. "Almost no LoRA" is not almost no change, and most of the slider's useful range does nothing.

Measurement

Rank-16 dora-rows adapter, ‖W − W0‖ / ‖W0‖ on the layer under test:

strength	before	after
0.0	0.00000%	0.00000%
1e-05	1.78760%	0.00000%
1e-03	1.78761%	0.00204%
0.01	1.78793%	0.04403%
0.1	1.82682%	0.41863%
0.5	2.59486%	2.08700%
1.0	4.16283%	4.16283%

Before: 43% of the total effect was already present at strength 1e-5, and the first 0.1 of the slider moved almost nothing. After: near-linear.

Fix

_blend_magnitude interpolates each learned magnitude toward the base weight's corresponding norm. Applied to all four magnitude-bearing adapters: dora, dora-xs, bora, bora-xs. The last two interpolate row and column magnitudes separately, since at strength 0 the row stage must reproduce W0 before the column stage can.

Compatibility

• strength 1.0 is bit-identical to before (the table's last row), so trained adapters and any strength-1 inference are unaffected.
• _blend_magnitude returns early at strength 1.0, so the common path — all of training — does not pay for the extra base-norm reduction.
• Anyone currently using a strength slider with a DoRA adapter will see different output at strengths other than 0 and 1. That is the point of the change, but it is a behaviour change and worth noting in release notes.

[Cortexelus]

Tested

Ran a full suite against this branch and against all three LoRA PRs merged together (they conflict in model.py — _blend_magnitude and _delta/baked_vnorm_row land at the same insertion point; both sides are purely additive, so the resolution is to keep both).

The discontinuity this fixes, measured on the one conditioner layer a real adapter touches (conditioners.seconds_total.embedder.embedding.1), as ‖W − W₀‖/‖W₀‖:

strength	before	after
0	0%	0%
1e-5	1.78760%	0.00000%
1.0	4.16283%	4.16283%

Full slider sweep (0 → 10, both a DiT layer and the conditioner layer):

0.0:0.00%  1e-05:0.00%  0.001:0.00%  0.1:0.42%  0.25:1.04%
0.5:2.09%  1.0:4.16%    2.0:8.27%    5.0:20.14%  10.0:37.81%

Asserted: exact base at strength 0, monotonic across the whole range, near-linear away from 0 (max/min slope ratio 1.006), amplification above 1.0 still usable to the slider's max of 10, round-trip back to 1.0 bit-exact, and strength 1.0 unchanged from the pre-PR value — so no existing output moves.

Also confirmed the wrapper's set_lora_strength dispatches to both model.model and model.conditioner, which is what makes the conditioner layer reachable by the slider at all.

54 assertions across the strength and gradio groups, all passing.