Use lazy tensor

[justinchuby]

• Use LazyTensor to avoid having the store to disk.
• Load model in native dtype with torch_dtype="auto" to avoid the cast overhead when loading the model.
• The end result is 6.7 GB of peak memory usage (75% reduction) w/o needing temp disk space & 28% reduction in export time
• Memory needed for a model is (size of the weights + max(size of individual weight) * 3)

With a bigger model (google/gemma-3-12b-it) (model size 22.7 GB):

Time: 0:14:38 -> 0:01:57 (87% reduction) Memory: 190.5 -> 35.3 (81% reduction)

(left: current main, right: with IR and lazy tensor)

Compare with gguf conversion tool