Add inference_speed.zip (repack/rename of scaling dataset)

Adds `inference_speed.zip` — the size-stratified protein dataset used by the new `inference_speed` benchmark, which replaces the old `scaling` benchmark (instadeepai/mlipaudit PR #148).

It is a repack of `scaling.zip`: identical 13 `.xyz` systems (71–6713 atoms), but the internal folder is renamed `scaling/` -> `inference_speed/` because the benchmark extracts the archive into `./data` and then reads from `./data/inference_speed/`. macOS `__MACOSX/` cruft was dropped.

`scaling.zip` is intentionally **kept** for now: the code PR is not yet merged, and released 0.1.x versions still ship the `scaling` benchmark and fetch `scaling.zip` at runtime. It can be removed in a follow-up once the code PR merges and old versions are retired.

inference_speed.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:709432de6c8aec43f961cc3bd1695188ab956375175e5e9a4f33460c77a0a9dc
+size 2197098