fix(patching): keep overlap-blend reassembly in the patch dtype (halve peak memory)

[vboussot]

Summary

Accumulator.assemble (overlap-blend reassembly) allocated weight_sum as a default float32
tensor. The final normalization result = result / weight_sum then promoted the whole
(channels × volume) accumulator to float32 even when the patches were float16 (as they are for
autocast/fp16 inference, where ModelComposite casts outputs to fp16).

For large multi-class reassemblies this silently doubled the peak memory of the assembled volume.
Concretely, TotalSegmentator's 118-class whole-body output at inference has an assembled tensor of
shape (118, D, H, W) — for total-3mm that is ~3.1e9 elements = 12.5 GB in fp32 vs 6.3 GB in fp16;
for the full total at 1.5 mm it is ~8× larger and was OOM-ing.

The fix

Allocate weight_sum with dtype=result.dtype, so result / weight_sum stays in the patch dtype and
the accumulator is never promoted to float32.

Effect (measured on TotalSegmentator-KonfAI:total-3mm, autocast fp16)

• Assembled volume dtype: float32 → float16; peak of that tensor halved (12.5 GB → 6.3 GB).
• Reassembly time: 17.65 s → 14.36 s.
• Output: 27 differing voxels / 15,068,484 (0.0000018%) vs the prior float32 path — fp16 rounding at
  class-boundary ties. The pipeline is already fp16 end-to-end (composite casts to fp16 before
  accumulation), so the float32 division was a spurious upcast, not a source of truth.

Test plan

• New test_blended_reassembly_preserves_patch_dtype: a multi-channel float16 Cosinus-blended
  reassembly must stay float16.
• Existing blend/partition-of-unity tests (test_patch_overlap_border, test_patching) and
  test_predictor_memory still pass (17 passed).

Independent of #26 (device placement); together they let TotalSegmentator run without crashing or OOM.