Revert #907 + #913: restore pre-#907 hf-cache rclone config

[huydhn]

Reverts #913 and #907, returning the hf-cache rclone mount to its pre-#907 config (no --use-mmap, no --buffer-size 4M — rclone defaults + the per-tier GOMEMLIMIT). This is the same config meta-prod-aws-ue1 runs today with zero small-tier OOMs.

#907 added --buffer-size 4M --use-mmap and #913 later dropped only --use-mmap; post-deploy verification on ue2 showed small-tier (256Mi CPU / 512Mi gpu1) OOMs continued regardless, so neither flag was the driver. The real cause is that rclone mount memory is tiered by GPU count while per-node RSS scales with pod density — ue2 packs runners onto 48xl/metal nodes, overwhelming the single 256Mi per-node mount. Reverting to the known-good baseline while the tiering fix is developed separately.

[jeanschmidt]

I believe we should keep the -buffer-size 4M anyways

[huydhn]

Superseding this in favor of #917, which keeps a per-tier --buffer-size (CPU 0 / GPU 4M) rather than reverting to rclone defaults. Post-deploy evidence showed the OOM driver is node pod-density on the small tiers, not the mmap/buffer flags per se, so a full revert doesn't address it.