Optimize dynamic initialization for LRE workers

[aaronmondal]

We currently evaluate the correct toolchain identifier on-the-fly:

nativelink/kubernetes/overlays/lre/worker-lre-cc.yaml

Lines 25 to 30 in 481226b

	args:
	- |
	git config --global --add safe.directory "*"
	NATIVELINK_WORKER_PLATFORM=docker://lre-cc:$(nix eval /mnt/src_root#lre-cc.imageTag --raw) &&
	printf '#!/bin/sh\nexport NATIVELINK_WORKER_PLATFORM=%s\nexec "$@"' "$NATIVELINK_WORKER_PLATFORM" > /entrypoint/entrypoint.sh &&
	chmod +x /entrypoint/entrypoint.sh

The "on-demand" evaluation was a straightforward approach to achieve strong reproducibility guarantees and to support running workers with different tags side-by-side. However, running the identifier evaluation during initializaion means that it slows down autoscaled workers since all of them have to run through the eval step.

There are many ways to improve the performance here, but we need to be careful not to impact the reproducibility guarantees of the existing setup. We also need to evaluate what the right place to fix this actually is.