hooks/pyramid_attention_broadcast: fix redundant recompute at iteration 0 and free stale cache when outside timestep range

[GitGlimpse895]

What does this PR do?

Fixes two bugs in PyramidAttentionBroadcastHook.new_forward:

1. Redundant iteration == 0 condition — self.state.cache is None already
   covers the first-call case after every reset_state, making the extra guard
   dead code that creates a misleading impression of two independent invariants.

2. Stale cache leaking GPU VRAM — when outside the active timestep range,
   the hook was still writing self.state.cache = output, holding a full
   hidden-state activation tensor on GPU until the next generation's
   reset_state call. For video transformers with dozens of PAB-hooked layers
   this accumulates hundreds of MBs of unreleased VRAM. The fix sets
   self.state.cache = None immediately when outside the range.

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum?
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@yiyixuxu @sayakpaul @DN6