Update llama3 8b fp16 mlir

[IanWood1]

Export command:

python3 -m sharktank.examples.export_paged_llm_v1 \
	--irpa-file=/shark-dev/data/llama3.1/weights/8b/fp16/llama3.1_8b_fp16.irpa \
	--output-mlir=model.mlir --output-config=/dev/null --bs-prefill=4 \
	--bs-decode=4  --attention-dtype=float16 --activation-dtype=float16 \
	--use-attention-mask --use-hf --kv-cache-dtype=float16

[kuhar]

Do we know what changed?

(Maybe should be fine to update the IR regardless?)

[IanWood1]

Do we know what changed?

(Maybe should be fine to update the IR regardless?)

Looks like there are some reshapes that aren't getting folded causing iree_linalg_ext.gather to fail to fuse. llvm/llvm-project#142827 fixes it for llama & mistral. With the change, the before vs after decode times are about equal.

[kuhar]

Looks like there are some reshapes that aren't getting folded causing iree_linalg_ext.gather to fail to fuse.

@Groverkss

[IanWood1]

Updating both instead at #16