fix(distributed): avoid duplicate FSDP2 prefetch all-gathers (#3411)

* fix(distributed): use non-reentrant HF checkpointing

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* test(distributed): run FSDP2 prefetch regression in CI

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* test(distributed): cover HF checkpointing in robustness CI

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* test(distributed): preserve Llama robustness coverage

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* test(ci): calibrate Llama cross-TP KL tolerance

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

---------

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>