[docs] attn_only: note it only saves attention for dispatcher-visible backends

The MUST_SAVE policy runs via __torch_dispatch__, so it only sees the
attention forward when it is a registered op (torch SDPA, FA4/cute). The
flash_attn-library FA2/FA3 training forward routes to the raw external func
(opaque to SAC), so its attention is recomputed — correct, but no speedup
over FULL for that backend. Document the caveat; no behavior change.

Author: rich7420

fastvideo/training/activation_checkpoint.py

@@ -130,7 +130,11 @@ def _apply_activation_checkpointing_attn_only(module: torch.nn.Module) -> torch.
     intermediates are huge but cheap to recompute — saving them is what makes
     the stock 'ops' mode OOM). Eliminates the attention forward recompute that
     FULL mode pays (flash_fwd runs 2x under FULL) while staying within memory.
-    Orthogonal to torch.compile."""
+    Orthogonal to torch.compile.
+
+    Backend caveat: only saves the attention forward for dispatcher-visible ops
+    (SDPA, FA4/cute); the flash_attn-lib FA2/FA3 forward is opaque to SAC and is
+    recomputed (correct, no speedup)."""
     from torch.utils.checkpoint import (CheckpointPolicy, create_selective_checkpoint_contexts)
 
     def _attn_only_policy(ctx, func, *args, **kwargs):